diff --git a/MANIFEST.in b/MANIFEST.in
index 35f8afd0..c430037e 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -65,4 +65,8 @@ include web/assets/tour/c8-s1.mp4
 include web/assets/tour/c8-s1.webp
 include web/assets/tour/c9-s1.mp4
 include web/assets/tour/c9-s1.webp
+include web/assets/tour/c10-person.webp
+include web/assets/tour/c10-s1.mp4
+include web/assets/tour/c10-s1.webp
+include web/assets/tour/c10-source.webp
 include web/assets/tour/fox-sheet.webp
diff --git a/__init__.py b/__init__.py
index 69298cab..886ebffe 100644
--- a/__init__.py
+++ b/__init__.py
@@ -61,4 +61,6 @@ async def comfy_entrypoint():
         logging.warning('[FreeVideo] Agent routes not registered: %s', error)
     from .freevideo_engine.comfy_nodes import FreeVideoExtension, register_prompt_cleanup
     register_prompt_cleanup()
+    from .freevideo_engine.comfy_nodes import schedule_startup_prefetch
+    schedule_startup_prefetch()
     return FreeVideoExtension()
diff --git "a/docs/zh-CN/\344\275\277\347\224\250\350\257\264\346\230\216.md" "b/docs/zh-CN/\344\275\277\347\224\250\350\257\264\346\230\216.md"
index 53e0c03e..76a1d8bc 100644
--- "a/docs/zh-CN/\344\275\277\347\224\250\350\257\264\346\230\216.md"
+++ "b/docs/zh-CN/\344\275\277\347\224\250\350\257\264\346\230\216.md"
@@ -1,18 +1,28 @@
 # 使用说明
 
-这一页说明创作面板里三件常被问到的事：用参考图生成视频、首帧尾帧和参考图的区别、做更长的视频。
+这一页说明创作面板里常用的操作：用参考图生成视频、替换人物、区分首帧尾帧和参考图、筛选作品，以及做更长的视频。
 
 ## 参考图生成视频
 
 1. 打开创作面板，展开“素材”。
 2. 点“上传文件”，选择图片、视频或音频。上传的每张图片可以指定用途：首帧、尾帧或参考。要让画面里出现这张图里的人物、物品或场景，选“参考”。
 3. 在描述中输入 `@`，从列表里选一张素材，描述里会出现 `<Picture 1>` 这样的标记。视频和音频素材对应 `<Video 1>`、`<Audio 1>`，各类素材分别编号。
-4. 在描述中写明这张图在画面中的作用，例如“`<Picture 1>` 里的女孩在海边奔跑”，然后点“生成视频”。
+4. 音频卡片上选“原样使用”时，成片直接使用这段声音，时长随音频而定；一次只能有一段音频原样使用。
+5. 在描述中写明这张图在画面中的作用，例如“`<Picture 1>` 里的女孩在海边奔跑”，然后点“生成视频”。
 
 素材可以暂时关闭而不删除：取消勾选的素材不参与这次生成，其余素材的编号会随之重排，描述里的标记会同步更新。
 
 参考图生成视频目前仍是实验能力。能够运行不等于每次都能保持人物外观一致，清晰的正面图效果更稳定。
 
+## 替换人物
+
+1. 展开创作面板的“素材”，打开“替换人物”。
+2. 在“源视频”中点“添加视频”，在“人物图”中点“添加图片”。
+3. 需要源视频的声音时，勾选“保留源视频声音”。
+4. 描述可以留空，也可以补充画面要求；然后点“生成视频”。
+
+替换人物不能与首帧、尾帧同时使用。首次使用时会自动下载替换人物模型（约 80 MB），面板中显示进度，下载完成后才能生成。源视频至少 1 秒；源视频没有声音时，不能保留声音。
+
 ## 首帧、尾帧和参考的区别
 
 | 用途 | 作用 | 适合 |
@@ -25,6 +35,10 @@
 - 首帧、尾帧暂不能与参考图、参考视频同时使用；同时给了两类，先保留其中一类。
 - 参考图保持原有比例，按输出像素面积缩小。参考视频只取与成片等长的一段，最长 15 秒，生成用时约为只用图片时的两倍。
 
+## 作品库和收藏
+
+在作品详情中点爱心图标收藏；列表上方选“收藏”只看收藏的作品，选“全部”看所有作品。
+
 ## 在描述中用 @ 引用素材
 
 - 输入 `@` 后会列出当前可用的素材，选中后插入对应的标记。
@@ -32,6 +46,10 @@
 - 编号顺序是：先是上传的素材，再是从上到下连接的图片，最后是音频。参考视频的声音也占一个音频编号。
 - 删除或关闭某张素材后，描述里指向它的标记会改成“[素材已移除]”，其余标记按新编号更新。
 
+## 出错了怎么办
+
+全部下载源都失败时，错误卡片会逐个列出下载源和失败原因。
+
 ## 用一句话做白模
 
 未连接外部 Agent 时，可开启聊天框中的白模按钮，用一句话描述镜头。有合适的现成场景时，会先生成白模供您确认。
@@ -42,7 +60,7 @@
 
 1. **直接把时长填长。** 超过一次能生成的长度时，FreeVideo 自动分成几段依次生成，再接成一条视频。
 2. **接着生成。** 在生成好的视频结果卡上点“接着生成”，继续下一段；也可在“剪辑视图”中逐段修改。适合边看边调。
-3. **用 AI 全自动。** 打开“AI 全自动”，把想法交给 Agent，它会按镜头拆分、逐镜生成并拼接。耗时随镜头数和每镜时长而定，每镜按普通生成计时。
+3. **用 AI 全自动。** 打开“AI 全自动”，把想法交给 Agent，它会按镜头拆分、逐镜生成并拼接。耗时随镜头数和每镜时长而定，每镜按普通生成计时。在对话列表中点对话旁的“更多”，选“重命名”可修改对话名称；Agent 开拍整部片前、重拍耗时较长时是否先询问，可在“设置”的“AI Agent”中选择。
 
 ### 在剪辑视图里加参考图
 
@@ -51,6 +69,7 @@
 ### 限制
 
 - 两种做法一样，一条视频最多 8 段（第一段加 7 次接着生成），有短于 10 秒的段时最多 4 段；创作面板里时长最多填 60 秒。接续的每一段至少 5 秒，每段都是 10 秒以上时整条约 110 秒。剪辑视图里逐段生成同样适用。段数越多，衔接处越容易出现变化，适合切换镜头处，可直接切镜头，用对应的参考图开始新的一段。
+- 每张显卡同一时间处理一个请求，其余请求在队列中等待。
 - 接着生成使用轻量档，需要 NVIDIA 显卡或 Mac，也不能和“继续二采”同时使用。
 - 单次生成的最长时长随画幅和显卡而不同，在创作面板的“可生成范围”中可查看各尺寸的最长时长和预计用时。
 - 分段生成同样需要 NVIDIA 显卡或 Mac。
diff --git a/freevideo_engine/adaptive.py b/freevideo_engine/adaptive.py
index e5355d6e..8a371cd7 100644
--- a/freevideo_engine/adaptive.py
+++ b/freevideo_engine/adaptive.py
@@ -41,6 +41,37 @@ def commit_exhausted(error):
     return found(error)
 
 
+def commit_limited_allocation(error, failure):
+    """A driver-level CUDA allocation failure on Windows with VRAM to spare.
+
+    WDDM charges device allocations to the system commit limit. When that limit
+    is spent the driver refuses to map more device memory although the card has
+    room: the error is the bare "CUDA error: out of memory" of the runtime, not
+    the caching allocator's "CUDA out of memory. Tried to allocate", and the
+    device is not near its budget. A worker that keeps serving afterwards
+    still holds the commit it charged (about 6 GiB private commit after a
+    cleaned-up failure, against 0.03 GiB in a new process) and its device
+    mappings are inconsistent: the retry in it failed with "resource already
+    mapped" and the worker could not be stopped. WDDM non-local exhaustion is a
+    separate case with its own recovery (``nonlocal_exhausted``).
+    """
+    from .system import windows
+    if not windows() or failure.get('kind') != 'gpu_oom' or nonlocal_exhausted(failure):
+        return False
+    gpu = failure.get('gpu') if isinstance(failure.get('gpu'), dict) else {}
+    free = gpu.get('free_bytes')
+    if type(free) is not int or free < 2 * 2**30:
+        return False
+    seen = set()
+    while error is not None and id(error) not in seen:
+        seen.add(id(error))
+        text = str(error).lower()
+        if 'cuda error: out of memory' in text and 'cuda out of memory' not in text:
+            return True
+        error = error.__cause__ or error.__context__
+    return False
+
+
 # A CUDA error that may leave the device itself unusable; never retried.
 DEVICE_STATE_ERRORS = ('cuda error: unknown error', 'device is lost', 'device lost', 'gpu is lost',
     'illegal memory access', 'device-side assert', 'unspecified launch failure', 'driver shutting down',
diff --git a/freevideo_engine/agent/cast.py b/freevideo_engine/agent/cast.py
index 948f2f3d..51fa615a 100644
--- a/freevideo_engine/agent/cast.py
+++ b/freevideo_engine/agent/cast.py
@@ -679,7 +679,11 @@ def generate(self, identity, value=None, *, by='builtin', ref=None, revision=Non
             if value['face'] and card['kind'] != 'character':
                 raise spec.SpecError([schema.diagnostic('/face', 'false for a setting card', True)])
             ready = value.get('source') != 'video' and getattr(self.jobs.port, 'image_ready', lambda: False)()
-            source = value.get('source', 'image' if ready else 'video')
+            from ..image_gen.loras import validate_selection
+            loras = validate_selection(value.get('loras', []))
+            source = value.get('source', 'image' if ready or loras else 'video')
+            if loras and source != 'image':
+                spec.fail('/loras', 'LoRAs for reference pictures require source=image')
             # An explicit image request goes through the image service even when
             # its pack is absent: that failure owns a diagnostic report and a
             # durable job/card, just like a picture requested directly.
@@ -689,7 +693,7 @@ def generate(self, identity, value=None, *, by='builtin', ref=None, revision=Non
                 from ..image_gen.cast_views import VIEWS, SETTING_VIEWS
                 # A picture takes seconds, so a character always gets every view, the face close-up included.
                 views = list(SETTING_VIEWS if card['kind'] == 'setting' else VIEWS)
-                items = self._image_items(card, views[:1], seed)
+                items = self._image_items(card, views[:1], seed, loras=loras)
             else:
                 parts = [('turn', SCENE if card['kind'] == 'setting' else TURN, False)]
                 if value['face']:
@@ -708,12 +712,14 @@ def generate(self, identity, value=None, *, by='builtin', ref=None, revision=Non
                 card['_generating'].update(source=source, stage='front', views=views[1:],
                                             job_views={jobs[0]: views[0]}, name=card['name'],
                                             description=card['description'])
+                if loras:
+                    card['_generating']['loras'] = loras
             result = dict(job=jobs[0], jobs=jobs, card=view(card))
             self._remember(record, ref, digest, result, client)
             self._save(record)
             return result
 
-    def _image_items(self, card, views, seed, front=None, from_photo=False):
+    def _image_items(self, card, views, seed, front=None, from_photo=False, loras=None):
         from ..image_gen.cast_views import size, uses_front, view_prompt, photo_prompt
         # A picture somebody put on the card (the user, or an agent passing on the user's photo) is the person.
         photo = next((row['media'] for row in card.get('images', []) if not row.get('generated')), None)
@@ -726,7 +732,8 @@ def _image_items(self, card, views, seed, front=None, from_photo=False):
                       view_prompt(dict(card, photo=True) if from_photo else card, part))
             items.append(dict(kind='image', prompt=prompt, width=width, height=height, seed=seed,
                               images=images, label=('Cast · %s · %s' % (card['name'], part))[:120],
-                              metadata=dict(cast_card=card['id'], cast_part=part)))
+                              metadata=dict(cast_card=card['id'], cast_part=part),
+                              **({'loras': loras} if loras else {})))
         return items
 
     def _image_views(self, identity, generating, hidden, rows):
@@ -742,7 +749,7 @@ def _image_views(self, identity, generating, hidden, rows):
             card = record['card']
             basis = dict(id=identity, kind=hidden['kind'], name=hidden['name'], description=hidden['description'])
             items = self._image_items(basis, hidden['views'], hidden['seed'], dict(file=front['result']['file']),
-                                      from_photo=bool(front['spec'].get('images')))
+                                      from_photo=bool(front['spec'].get('images')), loras=hidden.get('loras'))
             jobs, _ = self.jobs.submit_request(dict(jobs=items, ref='cast-' + hidden['ref'] + '-views'),
                 client='cast:' + identity, parent=dict(cast=identity, ref=hidden['ref']))
             card['generating']['jobs'].extend(row['id'] for row in jobs)
diff --git a/freevideo_engine/agent/comfy.py b/freevideo_engine/agent/comfy.py
index 77fbddd2..d9afb865 100644
--- a/freevideo_engine/agent/comfy.py
+++ b/freevideo_engine/agent/comfy.py
@@ -322,7 +322,11 @@ def vlm_cancel(self, key):
     # Local image generation ----------------------------------------------------------------------------
     def _image(self):
         from ..image_gen.service import get_service
-        return get_service(self.server)
+        service = get_service(self.server)
+        if (service is not None and getattr(service, 'folders', None) is None
+                and getattr(self, 'folders', None) is not None):
+            service.folders = self.folders
+        return service
 
     @property
     def image(self):
diff --git a/freevideo_engine/agent/errors.py b/freevideo_engine/agent/errors.py
index 2ddc5bf8..28ff01d3 100644
--- a/freevideo_engine/agent/errors.py
+++ b/freevideo_engine/agent/errors.py
@@ -7,6 +7,7 @@
 import math
 import re
 import uuid
+from ..image_gen.loras import ERRORS as LORA_ERRORS, MESSAGE_EN as LORA_MESSAGE_EN, MESSAGE_ZH as LORA_MESSAGE_ZH
 
 REQUEST_ID = ContextVar('agent_request_id', default=None)
 # The work session a token request belongs to (sessions.py); jobs and productions record it.
@@ -155,13 +156,14 @@ def image_failure(code, report_id=None):
     en, zh = IMAGE_REASONS.get(code, IMAGE_REASONS['invalid_request'] if code.startswith('invalid_')
                              or code == 'too_many_images' else IMAGE_REASONS['generation_failed'])
     valid = isinstance(report_id, str) and re.fullmatch(r'[0-9a-f]{32}', report_id)
-    return dict(reason=dict(en=en, zh=zh), report_id=report_id if valid and code not in (
+    return dict(reason=dict(en=en, zh=zh), report_id=report_id if valid and code not in LORA_ERRORS and code not in (
         'busy', 'cancelled', 'invalid_request', 'invalid_media', 'invalid_prompt', 'invalid_dimensions',
         'invalid_seed', 'too_many_images', 'disk_full') else None)
 
 
 # Recovery is a function of the committed task and phase, not just an error code.
 IMAGE_ERRORS = {
+    **LORA_MESSAGE_EN,
     'invalid_request': 'The picture request contains invalid fields.',
     'invalid_prompt': 'The picture prompt must contain text and be at most 2000 characters.',
     'invalid_dimensions': 'Picture width and height must be integers from 256 to 2048.',
@@ -183,6 +185,7 @@ def image_failure(code, report_id=None):
     'disk_full': IMAGE_REASONS['disk_full'][0],
     'cancelled': 'The picture job was cancelled.',
 }
+IMAGE_REASONS.update({code: (LORA_MESSAGE_EN[code], LORA_MESSAGE_ZH[code]) for code in LORA_ERRORS})
 for _code, _message in IMAGE_ERRORS.items():
     if _code not in ERRORS:
         ERRORS[_code] = (503 if _code in ('pack_missing', 'closed', 'image_unavailable') else 422,
@@ -267,6 +270,7 @@ def safe_details(value):
 # One public sentence shared by conversation cards, HTTP and MCP. Backend text
 # is classified, never echoed: it can contain prompts, filenames or credentials.
 FAILURE_REASONS_ZH = {
+    **LORA_MESSAGE_ZH,
     'lan_unavailable': '局域网设置需要启动器安装记录',
     'out_of_memory': '生成时显存或内存不足',
     'insufficient_memory': '可用显存或内存不够生成这段内容',
@@ -316,6 +320,7 @@ def safe_details(value):
     'internal_error': '没有完成',
 }
 FAILURE_REASONS_EN = {
+    **LORA_MESSAGE_EN,
     'lan_unavailable': 'Phone access requires FreeVideo installed with the launcher',
     'out_of_memory': 'Graphics or system memory ran out during generation',
     'insufficient_memory': 'There is not enough free graphics or system memory for this generation',
diff --git a/freevideo_engine/agent/jobs.py b/freevideo_engine/agent/jobs.py
index 12436623..75130776 100644
--- a/freevideo_engine/agent/jobs.py
+++ b/freevideo_engine/agent/jobs.py
@@ -1999,7 +1999,7 @@ def wait_image(self, row, code, *, waiting_for=None):
     def start_image(self, row):
         spec = row['spec']
         row['_phase'] = 'preparing images'
-        value = {k: spec[k] for k in ('width', 'height', 'seed') if k in spec}
+        value = {k: spec[k] for k in ('width', 'height', 'seed', 'loras') if k in spec}
         if spec.get('people'):
             # A first frame: the people's pictures at the asked places, in the scene shown or described.
             action = 'compose'
diff --git a/freevideo_engine/agent/schema.py b/freevideo_engine/agent/schema.py
index 8f37c4fd..ca82a58d 100644
--- a/freevideo_engine/agent/schema.py
+++ b/freevideo_engine/agent/schema.py
@@ -245,12 +245,21 @@ def vlm(*, local=False):
         ('kind', 'questions'), description='x-job-kind also checks the sum of images and sampled video frames across questions (at most 24).', **{'x-job-kind': 'vlm'})
 
 
+def image_loras():
+    from ..image_gen.loras import MAX_LORAS, MAX_NAME, MIN_STRENGTH, MAX_STRENGTH
+    # The name's rules (extension, folders, control characters) are checked by validate_selection with its own messages.
+    return array(obj(dict(name=dict(type='string', minLength=1, maxLength=MAX_NAME), strength=number(MIN_STRENGTH, MAX_STRENGTH, default=1.),
+                          enabled=dict(type='boolean', default=True)), ('name',)), high=MAX_LORAS,
+                 **{'x-image-loras': True})
+
+
 def image(*, local=False):
     return obj(dict(kind={'const': 'image', 'type': 'string'},
         prompt=dict(type='string', minLength=1, maxLength=2000),
         width=number(256, 2048, integer=True), height=number(256, 2048, integer=True),
         aspect=enum(IMAGE_ASPECTS, description='Instead of width and height; 1:1 when all three are omitted.'),
         seed=nullable(number(-2**63, 2**64 - 1, integer=True)),
+        loras=image_loras(),
         images=array({'oneOf': [FILE, obj(dict(file=FILE), ('file',))]}, high=4, description='Pictures to change '
             'or combine: input files, such as result.file of an earlier picture job. The prompt calls them image 1, image 2.'),
         people=array(obj(dict(image=FILE, cast=CAST, name=string(40), x=number(0, 1), height=number(.05, 1),
@@ -613,7 +622,8 @@ def cast_schema(name, *, local=False):
         return obj(dict(revision=number(1, integer=True), ref=REF))
     if name == 'generate':
         return obj(dict(face=dict(type='boolean', default=False, description='Also make one face close-up (video; the image model always does).'),
-                        source=enum(('image', 'video'), description='Defaults to image when the local image model is ready, otherwise video.'),
+                        loras=image_loras(),
+                        source=enum(('image', 'video'), description='Defaults to image when LoRAs are selected or the local image model is ready, otherwise video.'),
                         replace=dict(type='boolean', default=False,
                                      description='Replace pictures the user did not add; default fills empty slots.'),
                         revision=number(1, integer=True), ref=REF))
@@ -1227,6 +1237,12 @@ def cross_issues(schema, value, path):
     errors = []
     def bad(field, allowed, got=None):
         errors.append(diagnostic(path + '/' + field, allowed, got))
+    if schema.get('x-image-loras'):
+        from ..image_gen.loras import MESSAGE_ZH, validate_selection
+        try:
+            validate_selection(value)
+        except ValueError as error:
+            errors.append(diagnostic(path, MESSAGE_ZH[str(error)], None))
     flags = schema.get('x-media')
     if flags is not None:
         item = {'file': value} if isinstance(value, str) else value
@@ -1551,7 +1567,7 @@ def openapi():
     description='Work session; omit for latest.')
 
 
-SHAPE_MIN_BYTES = 400
+SHAPE_MIN_BYTES = 200
 RECIPE_TOOL = 'freevideo_edit_image'
 
 
diff --git a/freevideo_engine/agent/spec.py b/freevideo_engine/agent/spec.py
index e70c8b95..46846f8d 100644
--- a/freevideo_engine/agent/spec.py
+++ b/freevideo_engine/agent/spec.py
@@ -271,6 +271,10 @@ def normalize(value, *, port=None):
             result['questions'].append(item)
         result.update({k: value[k] for k in ('format', 'max_tokens', 'system') if k in value})
     elif kind == 'image':
+        from ..image_gen.loras import validate_selection
+        loras = validate_selection(original.get('loras', []))
+        if loras:
+            result['loras'] = loras
         width, height = ((value['width'], value['height']) if 'width' in value else
                          contract.IMAGE_ASPECTS[value.get('aspect', '1:1')])
         result.update(prompt=value.get('prompt', ''), width=width, height=height,
diff --git a/freevideo_engine/bootstrap.py b/freevideo_engine/bootstrap.py
index f477ceea..c238093f 100644
--- a/freevideo_engine/bootstrap.py
+++ b/freevideo_engine/bootstrap.py
@@ -158,7 +158,7 @@ def reinstall_download_bytes(dependencies, system, *, root=None, frontend_python
     Windows wheel (torch_download), elsewhere the package's installed size, an
     upper bound for its wheel. Marks each engine environment's damaged packages."""
     from .environments import ENVIRONMENTS
-    from .runtime_libraries import damaged, incomplete, installed_bytes
+    from .runtime_libraries import damaged, incomplete, installed_bytes, missing_packages
     from .torch_download import select
     total = 0
     seen = set()
@@ -173,6 +173,9 @@ def reinstall_download_bytes(dependencies, system, *, root=None, frontend_python
         row['damaged'] = broken = damaged(row['python'])
         if frontend:
             broken = sorted(set(broken + incomplete(row['python'])))
+        else:
+            # Setup puts back pinned packages whose files are gone (Setup.repairing).
+            broken = sorted(set(broken + missing_packages(row['python'], constraints_file(name, system))))
         pins = ENVIRONMENTS.get(name, {})
         wheels = {w['package']: w for w in (select(pins['torch'], pins['cuda']) or [])} if system == 'Windows' and pins else {}
         for package in broken:
@@ -1628,10 +1631,11 @@ def check():
                     if install(uv, executable, role['torch'], role['cuda'], root=self.root,
                                networking=self.network, env=self.env, ui=self.ui,
                                run=lambda args, env: self.command(label + '-torch', args, env=env),
-                               constraints=['-c', constraints_file(label, self.system)], check=check):
+                               constraints=['-c', constraints_file(label, self.system)], check=check,
+                               reinstall=self.missing_files(executable, label, gpu=True, wheels=role)):
                         return
                 from .runtime_libraries import damaged, reinstall_arguments
-                fresh = self.check_runtime_libraries(executable, label)
+                fresh = sorted(set(self.check_runtime_libraries(executable, label)) | set(self.missing_files(executable, label, gpu=True)))
                 result = self.packages(label + '-torch', [*argv[:5], *reinstall_arguments(fresh), *argv[5:]], group)
                 # Files linked from a changed uv cache arrive with a new environment.
                 if not fresh:
@@ -1647,7 +1651,8 @@ def check():
                    '-c', constraints_file(name, self.system), 'pip', 'ninja', 'packaging', 'wheel', 'setuptools']
         if self.layout == 'unified':
             command += ['-r', SOURCE / 'constraints/encoder-runtime.txt']
-        task('runtime', lambda: self.packages(name + '-packages', command), (name + '-torch',), ('environment:' + name,))
+        task('runtime', lambda: self.packages(name + '-packages', self.repairing(command, python, name)),
+             (name + '-torch',), ('environment:' + name,))
         # Enqueue downloads as soon as their SDKs exist, before waiting for
         # source setup, encoder checks or acceleration compilation.
         task('models', lambda: self.command('models', [python, '-m', 'freevideo_engine.provision',
@@ -1657,8 +1662,9 @@ def check():
         encoder_after = 'runtime'
         if self.layout == 'dual':
             def encoder_packages():
-                self.packages('encoder-packages', [uv, 'pip', 'install', '--python', encoder_python,
-                    '-r', SOURCE / 'constraints/encoder-runtime.txt', '-c', constraints_file('encoder', self.system)])
+                self.packages('encoder-packages', self.repairing([uv, 'pip', 'install', '--python', encoder_python,
+                    '-r', SOURCE / 'constraints/encoder-runtime.txt', '-c', constraints_file('encoder', self.system)],
+                    encoder_python, 'encoder'))
                 self.packages('encoder-engine', [uv, 'pip', 'install', '--python', encoder_python, '--no-deps', '-e', SOURCE])
             task('encoder-runtime', encoder_packages, ('encoder-torch',), ('environment:encoder',))
             encoder_after = 'encoder-runtime'
@@ -1690,6 +1696,33 @@ def encoder_packages():
             tasks['models']['after'] = ('package-cache',)
         return tasks, pythons, comfy
 
+    def missing_files(self, python, role, *, gpu=None, wheels=None):
+        """Pinned packages of an environment whose files are gone (runtime_libraries.package_files).
+
+        gpu=True keeps the PyTorch family, which is installed from the GPU wheels (and only
+        those of `wheels`, an environment's spec, when given); gpu=False keeps the rest."""
+        from .runtime_libraries import gpu_family, missing_packages, normalize_name
+        try:
+            names = missing_packages(python, constraints_file(role, self.system))
+        except ValueError:
+            return []
+        if gpu is not None:
+            names = [name for name in names if gpu_family(name) == gpu]
+        if wheels is not None:
+            from .torch_download import select
+            catalog = {normalize_name(row['package']) for row in select(wheels['torch'], wheels['cuda']) or []}
+            names = [name for name in names if normalize_name(name) in catalog]
+        return names
+
+    def repairing(self, argv, python, role):
+        """An install command that also replaces the pinned packages whose files are gone.
+
+        Installing the pinned version again does nothing when uv finds it installed,
+        whatever state its files are in. The PyTorch family is replaced in its own step."""
+        from .runtime_libraries import reinstall_arguments
+        names = self.missing_files(python, role, gpu=False)
+        return [*argv[:5], *reinstall_arguments(names), *argv[5:]] if names else argv
+
     def check_runtime_libraries(self, python, role):
         from .runtime_libraries import damaged
         key = role + '-runtime-libraries'
diff --git a/freevideo_engine/comfy_nodes.py b/freevideo_engine/comfy_nodes.py
index 2ff56df0..93a12312 100644
--- a/freevideo_engine/comfy_nodes.py
+++ b/freevideo_engine/comfy_nodes.py
@@ -4,7 +4,7 @@
 
 from comfy_api.latest import ComfyExtension, InputImpl, io, ui
 
-from . import comfy_bridge, workflow_repairs
+from . import comfy_bridge, usage_habit, workflow_repairs
 from .comfy_continuation import node_frames, optional_text, part_steps, request_options
 from .default_prompt import DEFAULT_PROMPT
 from .original_audio import audio_use_option
@@ -56,6 +56,26 @@ def register_prompt_cleanup():
     server.add_on_prompt_handler(clean_prompt)
 
 
+def schedule_startup_prefetch(delay=15.):
+    """Once after the engine starts, read what the first request reads into the OS file cache, as far as
+    free memory goes. A request, or the busy queue, cancels it through the same stop as the idle read."""
+    from .encoder_prewarm import IDLE, startup_files
+    try:
+        from server import PromptServer
+        server = PromptServer.instance
+        root, machine = comfy_bridge.installation()
+        habit = usage_habit.load()
+        files = startup_files(machine, usage_habit.dominant(habit['recent'] if habit else None))
+        files = [item for item in files if item.is_file()]
+        if not files:
+            return False
+    except (ImportError, OSError, RuntimeError, ValueError, KeyError, TypeError, AttributeError):
+        return False  # Optional preparation cannot stop the engine from starting.
+    def busy():
+        return server is not None and server.prompt_queue.get_tasks_remaining() > 0
+    return IDLE.schedule(files, root / 'startup-prewarm.json', busy=busy, purpose='startup', delay=delay)
+
+
 def step_cache_option(value):
     """Bridge keyword for the Step cache widget; a workflow without it, with 'none' or with a value that is
     not an option adds nothing. Workflows saved before the widget existed hold '' in its position."""
@@ -238,6 +258,11 @@ def execute(cls, text, width, height, seconds, seed, first=None, last=None,
             stop_music_for_generation(PromptServer.instance)
         except Exception as error:
             logging.warning('[FreeVideo] Could not stop music before the video: %s', error)
+        try:   # the picture worker keeps the card longer only for people who mostly make pictures
+            from .image_gen.keep import note, path
+            note('video', path(comfy_bridge.installation_root()))
+        except Exception as error:
+            logging.warning('[FreeVideo] Could not record the video for the picture worker: %s', error)
         prewarm = {}
         result_reused = [False]
         result_report = [None]
@@ -397,8 +422,8 @@ def busy():
             gpu = comfy_bridge.installation()[1].get('gpu_uuid')
         except (OSError, ValueError, KeyError, TypeError):
             gpu = None
-        OWNER.finished(next_part=next_part, gpu=gpu, warm=True,
-                       agent=server is not None and str(getattr(server, 'client_id', '') or '').startswith('agent:'))
+        kept = OWNER.finished(next_part=next_part, gpu=gpu, warm=True,
+                              agent=server is not None and str(getattr(server, 'client_id', '') or '').startswith('agent:'))
         from .comfy_environment import executor_device
         if target and not result_reused[0] and not executor_device():
             client = server.client_id if server is not None else None
@@ -411,12 +436,19 @@ def warmed(value):
                     IDLE.schedule(path, output.with_suffix('.prewarm.json'),
                                   ram_budget=measured.get('profile', {}).get('inference_ram_budget_gb', 0) * 1e9,
                                   busy=busy, notify=warmed, purpose='upscale')
+                elif (isinstance(kept, dict) and kept.get('action') == 'kept' and next_part is None
+                        and usage_habit.dominant((usage_habit.load() or {}).get('recent')) == 'prompt'):
+                    # Prompts change often: place the encoder on the card while the kept worker
+                    # idles. It yields to any new task, and only what fits free memory is placed.
+                    root, machine = comfy_bridge.installation()
+                    environment = comfy_bridge.engine_environment(root, comfy_bridge.source_root())
+                    environment.update(FREEVIDEO_HOME=str(root), PYTHONPATH=str(comfy_bridge.source_root()))
+                    OWNER.warm_encoder(output, machine['python'], environment, busy=busy, notify=warmed)
                 else:
-                    # File cache only: a kept worker must not grow on the card while it waits.
                     IDLE.schedule(path, output.with_suffix('.prewarm.json'),
                                   ram_budget=measured.get('profile', {}).get('inference_ram_budget_gb', 0) * 1e9,
                                   busy=busy, notify=warmed)
-            except (OSError, RuntimeError):
+            except (OSError, RuntimeError, ValueError, KeyError, TypeError):
                 pass  # A saved video remains successful if idle warming cannot start.
         return io.NodeOutput(InputImpl.VideoFromFile(str(output)), report, ui=preview)
 
diff --git a/freevideo_engine/comfy_residency.py b/freevideo_engine/comfy_residency.py
index aa329952..d9dff576 100644
--- a/freevideo_engine/comfy_residency.py
+++ b/freevideo_engine/comfy_residency.py
@@ -76,6 +76,16 @@ def release_native(server, *, timeout=10.):
                 reserved_after_bytes=int(torch.cuda.memory_reserved(device)))
 
 
+def release_image_worker():
+    """Let an idle image worker give back the card; one that is making a picture is left alone."""
+    try:
+        from server import PromptServer
+        from .image_gen.service import release_idle_worker
+        release_idle_worker(PromptServer.instance)
+    except Exception:
+        pass  # Native models must load whether or not the image option is installed.
+
+
 def install(memory):
     if getattr(memory.free_memory, '_freevideo_resident_cooperation', False) is True:
         return
@@ -90,6 +100,8 @@ def free_memory(memory_required, device, *args, **kwargs):
                 OWNER.process is not None and OWNER.process.poll() is None):
             if device is None or memory_required > memory.get_free_memory(device):
                 OWNER.release('comfy_model')
+        if device is None or memory_required > memory.get_free_memory(device):
+            release_image_worker()
         return original(memory_required, device, *args, **kwargs)
 
     free_memory._freevideo_resident_cooperation = True
diff --git a/freevideo_engine/comfy_upload.py b/freevideo_engine/comfy_upload.py
index 5a724af7..aaaf0899 100644
--- a/freevideo_engine/comfy_upload.py
+++ b/freevideo_engine/comfy_upload.py
@@ -110,10 +110,18 @@ async def lora_check(request):
         from .lora_check import check_installed_file
         try:
             name = request.query.get('name')
-            path = resolve_loras(json.dumps([dict(name=name)]), folder_paths.get_filename_list('loras'),
-                                 lambda n: folder_paths.get_full_path_or_raise('loras', n))[0]['path']
-            result = await asyncio.to_thread(check_installed_file, path)
+            if request.query.get('kind') == 'image':
+                from .image_gen.loras import check_named_file
+                result = await asyncio.to_thread(check_named_file, name, folders=folder_paths)
+            else:
+                path = resolve_loras(json.dumps([dict(name=name)]), folder_paths.get_filename_list('loras'),
+                                     lambda n: folder_paths.get_full_path_or_raise('loras', n))[0]['path']
+                result = await asyncio.to_thread(check_installed_file, path)
         except (OSError, ValueError, TypeError) as error:
+            if request.query.get('kind') == 'image':
+                from .image_gen.loras import MESSAGE_ZH
+                return web.json_response({'error': str(error),
+                                          'message_zh': MESSAGE_ZH.get(str(error), MESSAGE_ZH['invalid_loras'])}, status=400)
             return web.json_response({'error': str(error)}, status=400)
         return web.json_response(result, headers={'Cache-Control': 'no-store'})
 
diff --git a/freevideo_engine/diagnostic_resources.py b/freevideo_engine/diagnostic_resources.py
index ea57cffa..1200afdc 100644
--- a/freevideo_engine/diagnostic_resources.py
+++ b/freevideo_engine/diagnostic_resources.py
@@ -143,7 +143,7 @@ def encoder_prewarm(value):
     for key in ('gpu_preload', 'worker_released', 'receipt_available'):
         if type(value.get(key)) is bool:
             result[key] = value[key]
-    if value.get('purpose') in ('encoder', 'upscale'):
+    if value.get('purpose') in ('encoder', 'upscale', 'startup'):
         result['purpose'] = value['purpose']
     # Why idle preparation stopped, as fixed words only: the receipt's reason
     # is free text and may name a path. A bare "failed" explained nothing.
diff --git a/freevideo_engine/encode_worker.py b/freevideo_engine/encode_worker.py
index fc3a534e..4ceba4ec 100644
--- a/freevideo_engine/encode_worker.py
+++ b/freevideo_engine/encode_worker.py
@@ -576,8 +576,8 @@ def attempt(index, failure):
             save(request['metrics'], {**loading, 'success': False, 'phase': 'keyframe_vae',
                                      'load_seconds': load_seconds, 'prompt_encode_seconds': prompt_encode_seconds})
         # Only the encoder is read; the decoder in the shared shards is skipped.
-        vae, _ = load_video_encoder(request['base'], before_upload=None if resident is None else
-                                    lambda model: resident.input_vae_room(model, dict(width=width, height=height)))
+        vae = (resident.input_video_encoder(request['base'], dict(width=width, height=height))
+               if resident is not None else load_video_encoder(request['base'])[0])
         torch.cuda.synchronize()
         keyframe_metrics['keyframe_vae_load_seconds'] = time.perf_counter() - vae_started
         vae_started = time.perf_counter()
diff --git a/freevideo_engine/encoder_prewarm.py b/freevideo_engine/encoder_prewarm.py
index 153bdc15..a2af164c 100644
--- a/freevideo_engine/encoder_prewarm.py
+++ b/freevideo_engine/encoder_prewarm.py
@@ -18,15 +18,23 @@
 CHUNK = 8 * 1024 * 1024
 
 
-def plan(size, available, *, ram_budget=None, complete=True):
+MINIMUM_PARTIAL = 256 * 1024 * 1024
+
+
+def plan(size, available, *, ram_budget=None, complete=True, partial=False):
+    """Choose how much of the files to read. A sequential read of more than the cache holds evicts its own
+    earlier pages, so a partial read stops at the memory that is really free instead of exceeding it."""
     reserve = 2 * GiB
     limit = max(0, available - reserve)
     if ram_budget is not None:
         limit = min(limit, max(0, int(ram_budget)))
-    return dict(enabled=bool(complete and size > 0 and size <= limit), bytes=size,
+    whole = bool(complete and size > 0 and size <= limit)
+    part = bool(partial and complete and size > limit >= MINIMUM_PARTIAL)
+    return dict(enabled=whole or part, partial=part, bytes=size, read_bytes=size if whole else (limit if part else 0),
                 memory_allowance_bytes=limit, system_reserve_bytes=reserve,
                 scope='Reclaimable OS file cache; no private model copy or GPU allocation',
-                reason=('Idle RAM covers the checkpoint plus system headroom' if complete and 0 < size <= limit
+                reason=('Idle RAM covers the checkpoint plus system headroom' if whole
+                        else 'Idle RAM covers the first part of the files, read in the order of use' if part
                         else 'Leave memory for active applications; encoder loads on demand'))
 
 
@@ -45,6 +53,16 @@ def warm_target(report):
     return ('encoder', encoder) if encoder else None
 
 
+def startup_files(machine, habit):
+    """Files the first request after an engine start reads first: a rerun reads the video model blocks,
+    any other habit, or none yet, starts with the text encoder."""
+    if habit == 'rerun':
+        files = sorted(item for item in Path(machine['checkpoint']).rglob('*') if item.is_file())
+        if files:
+            return files
+    return [Path(machine['encoder_model_root']) / 'text_encoders' / machine['encoder']]
+
+
 def available_memory():
     memory = system_memory()
     group = cgroup_memory()
@@ -67,12 +85,14 @@ def prefetch(path, selected, cancelled, *, busy=None, notify=None, availability=
         return result
     result['state'] = 'reading'
     buffer = bytearray(CHUNK)
+    view = memoryview(buffer)
     last = 0.
+    limit = selected.get('read_bytes', sum(item.stat().st_size for item in paths))
     for item in paths:
         initial = item.stat()
         remaining = initial.st_size
         with item.open('rb', buffering=0) as stream:
-            while remaining > 0:
+            while remaining > 0 and result['done_bytes'] < limit:
                 if cancelled.is_set() or busy and busy():
                     result.update(state='cancelled', reason='A new request takes priority')
                     break
@@ -80,7 +100,7 @@ def prefetch(path, selected, cancelled, *, busy=None, notify=None, availability=
                 if not complete or available < selected['system_reserve_bytes'] + CHUNK:
                     result.update(state='released', reason='Live RAM pressure; warmed pages remain reclaimable')
                     break
-                count = stream.readinto(buffer)
+                count = stream.readinto(view[:min(CHUNK, limit - result['done_bytes'])])
                 if not count:
                     raise ValueError('A model file changed during prefetch')
                 remaining -= count
@@ -92,6 +112,9 @@ def prefetch(path, selected, cancelled, *, busy=None, notify=None, availability=
                     last = elapsed
         if result['state'] != 'reading':
             break
+        if remaining > 0:
+            result.update(state='partial', reason='The first part of the files is read; the rest loads on demand')
+            break
         current = item.stat()
         if (current.st_size, current.st_mtime_ns) != (initial.st_size, initial.st_mtime_ns):
             raise ValueError('A model file changed during prefetch')
@@ -116,7 +139,7 @@ def stop(self):
         from .encoder_diagnostics import prewarm_receipt
         return prewarm_receipt(self.report, gpu=False)
 
-    def schedule(self, path, report, *, ram_budget=None, busy=None, notify=None, purpose='encoder'):
+    def schedule(self, path, report, *, ram_budget=None, busy=None, notify=None, purpose='encoder', delay=2.):
         """Warm the encoder checkpoint, or the files an upscale of a preview reads."""
         paths = [Path(item) for item in path] if isinstance(path, (list, tuple)) else [Path(path)]
         with self.guard:
@@ -130,11 +153,11 @@ def work():
                 result = {'state': 'cancelled', 'reason': 'A new request takes priority', 'purpose': purpose}
                 try:
                     # Return the completed video to the user before doing work.
-                    if cancel.wait(2.) or busy and busy():
+                    if cancel.wait(delay) or busy and busy():
                         return
                     available, complete = available_memory()
                     selected = plan(sum(item.stat().st_size for item in paths), available,
-                                    ram_budget=ram_budget, complete=complete)
+                                    ram_budget=ram_budget, complete=complete, partial=purpose == 'startup')
                     selected['purpose'] = purpose
                     if not selected['enabled'] and purpose == 'upscale':
                         selected['reason'] = 'Leave memory for active applications; the upscale reads its files on demand'
diff --git a/freevideo_engine/failure_details.py b/freevideo_engine/failure_details.py
index 7d8ec3d7..af9d23c8 100644
--- a/freevideo_engine/failure_details.py
+++ b/freevideo_engine/failure_details.py
@@ -756,7 +756,9 @@ def _launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_re
          ('Runtime files have been changed', '运行组件已被更改'),
          ('Click Repair to reinstall a complete copy of PyTorch. Models and videos are kept.',
           '请点“修复”，FreeVideo 会重新安装运行组件，模型和视频都会保留。')),
-        (r'ModuleNotFoundError|No module named|DLL load failed', 'dependencies', *DEPENDENCIES),
+        # runtime_libraries.summary names packages whose files are gone (a numpy extension that cannot load).
+        (r'ModuleNotFoundError|No module named|DLL load failed|Files of these Python packages are missing or have changed',
+         'dependencies', *DEPENDENCIES),
         (r'CUDA (?:error: )?out of memory|torch\.OutOfMemoryError|Insufficient currently available memory|crossed its RAM budget|paging file is too small|WinError 1455|os error 1455', 'memory',
          ('Memory is unavailable for this step', '这一步可用内存不足'),
          ('Close other memory-heavy applications and retry. Copy the details if it continues.', '关闭占用内存较多的程序后重试；仍失败时可复制详情反馈。')),
diff --git a/freevideo_engine/image_gen/compose.py b/freevideo_engine/image_gen/compose.py
index 0dc06538..9065270d 100644
--- a/freevideo_engine/image_gen/compose.py
+++ b/freevideo_engine/image_gen/compose.py
@@ -137,4 +137,5 @@ def load(name):
         load(name)
     if len(prompt) > 2000:
         raise ValueError('invalid_prompt')
-    return dict(prompt=prompt, width=width, height=height, seed=value.get('seed'), images=names)
+    return dict(prompt=prompt, width=width, height=height, seed=value.get('seed'), images=names,
+                **({'loras': value['loras']} if 'loras' in value else {}))
diff --git a/freevideo_engine/image_gen/failure.py b/freevideo_engine/image_gen/failure.py
index 57af20d7..08823270 100644
--- a/freevideo_engine/image_gen/failure.py
+++ b/freevideo_engine/image_gen/failure.py
@@ -13,16 +13,18 @@
 from ..exit_diagnostics import clean_reading, clean_sample
 from ..diagnostic_resources import commit_exhausted
 from ..failure_details import Redactor
+from .loras import ERRORS as LORA_ERRORS, MESSAGE_ZH as LORA_MESSAGE_ZH
 
 
 ERRORS = frozenset(('no_memory', 'small_memory', 'busy_video', 'busy', 'pack_missing', 'model_unreadable',
                     'device_unavailable', 'worker_exit', 'timeout', 'cancelled', 'disk_full',
-                    'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'generation_failed'))
+                    'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'generation_failed')) | LORA_ERRORS
 NO_REPORT = frozenset(('small_memory', 'busy', 'invalid_request', 'invalid_media', 'invalid_points', 'segment_not_installed', 'cancelled',
-                       'disk_full'))
+                       'disk_full')) | LORA_ERRORS
 STAGES = ('start', 'load', 'text', 'encode', 'denoise', 'decode')
 KINDS = ('generate', 'edit', 'cast', 'compose', 'download', 'install')
 MESSAGE_ZH = {
+    **LORA_MESSAGE_ZH,
     'no_memory': '显存不够，最小尺寸也没能生成。关掉其他占用显卡的程序后再试。',
     'small_memory': '这台 Mac 的内存不足 16 GB，无法生成图片。',
     'busy_video': '视频正在生成，完成后再试。',
diff --git a/freevideo_engine/image_gen/image_runner.py b/freevideo_engine/image_gen/image_runner.py
index e4e11cac..8957c7d6 100644
--- a/freevideo_engine/image_gen/image_runner.py
+++ b/freevideo_engine/image_gen/image_runner.py
@@ -293,7 +293,7 @@ def plan(self, width, height, reference_images=None):
             reference_pixels=reference_pixels, shape=self.shape)
 
     def generate(self, prompt, *, width, height, seed=None, reference_images=None, edit_mask=None,
-                 progress=None, cancel=None):
+                 loras=None, progress=None, cancel=None):
         plan = self.plan(width, height, reference_images)
         if self.full_card and not self._all_on_card(plan):
             self._to_streamed()
@@ -305,9 +305,10 @@ def generate(self, prompt, *, width, height, seed=None, reference_images=None, e
         self.te_stack.set_resident(plan.te_resident_layers)
         self.generator.vae_tile = plan.vae_tiling
         self.generator.vae_tile_size = plan.vae_tile_size
+        extra = dict(loras=loras) if loras else {}
         return self.generator.generate(prompt, width=width, height=height, seed=seed,
                                        reference_images=reference_images, edit_mask=edit_mask,
-                                       progress=progress, cancel=cancel)
+                                       progress=progress, cancel=cancel, **extra)
 
     def _all_on_card(self, plan):
         return (plan.feasible and _kept(plan) == len(self.dit_stack.layers)
diff --git a/freevideo_engine/image_gen/keep.py b/freevideo_engine/image_gen/keep.py
new file mode 100644
index 00000000..a00325bc
--- /dev/null
+++ b/freevideo_engine/image_gen/keep.py
@@ -0,0 +1,78 @@
+"""How long an idle image worker keeps the model on the card.
+
+A worker that made a picture waits IDLE_SECONDS for the next one. For someone who mostly makes
+pictures it waits KEEP_SECONDS, but only while the whole model sits on the card, and it lets go at
+once when another program starts taking card memory. Video always takes the card first.
+"""
+
+from .. import usage_habit
+
+IDLE_SECONDS = 45.
+KEEP_SECONDS = 600.
+# Operations looked at (two thirds of them must be pictures), and how few are enough to judge at all.
+WINDOW = 6
+MINIMUM = 3
+STORED = usage_habit.OPS_STORED
+# Card memory other programs may take while the worker waits before it lets go.
+YIELD_BYTES = 1 << 30
+POLL_SECONDS = .5
+
+
+def path(root=None):
+    """Where the habit is kept: the usage habit file under the installation directory."""
+    return usage_habit.path(root)
+
+
+def load(file=None):
+    """The last operations ('image' or 'video'), oldest first; empty when the file cannot be trusted."""
+    return usage_habit.ops(file)
+
+
+def note(kind, file=None):
+    """Remember one operation. Only the kind is stored, nothing from the request."""
+    usage_habit.note_op(kind, file)
+
+
+def mostly_images(file=None):
+    """True when most of the recent operations were pictures and the last one was too."""
+    recent = load(file)[-WINDOW:]
+    return len(recent) >= MINIMUM and recent[-1] == 'image' and recent.count('image') * 3 >= len(recent) * 2
+
+
+def keep_seconds(file=None):
+    """The idle limit the worker is asked to use for the request being sent."""
+    return KEEP_SECONDS if mostly_images(file) else IDLE_SECONDS
+
+
+def card_free_bytes(device):
+    """Free memory of the card the worker's model is on, or None when it cannot be read."""
+    torch = __import__('sys').modules.get('torch')
+    if torch is None or getattr(device, 'type', str(device).split(':')[0]) != 'cuda':
+        return None
+    try:
+        return int(torch.cuda.mem_get_info(device)[0])
+    except Exception:
+        return None
+
+
+class Crowding:
+    """True once other programs have taken YIELD_BYTES of card memory since the wait began.
+
+    The worker's own memory does not change while it waits, so a drop in free memory is theirs."""
+
+    def __init__(self, read, limit=YIELD_BYTES):
+        self.read = read
+        self.limit = limit
+        self.start = read()
+
+    def __bool__(self):
+        now = self.read()
+        return self.start is not None and now is not None and self.start - now >= self.limit
+
+
+def wait_seconds(requested, runner, idle=IDLE_SECONDS):
+    """Idle limit for the worker after a picture: the longer one only for a model whole on a readable card."""
+    if (type(requested) in (int, float) and requested > idle and getattr(runner, 'full_card', False) is True
+            and card_free_bytes(getattr(runner, 'device', None)) is not None):
+        return min(float(requested), KEEP_SECONDS)
+    return idle
diff --git a/freevideo_engine/image_gen/loras.py b/freevideo_engine/image_gen/loras.py
new file mode 100644
index 00000000..63023ba0
--- /dev/null
+++ b/freevideo_engine/image_gen/loras.py
@@ -0,0 +1,111 @@
+"""Shared image LoRA admission; names resolve only through ComfyUI's libraries."""
+import json
+import math
+from pathlib import Path
+
+MAX_LORAS = 8
+MAX_NAME = 512
+MIN_STRENGTH, MAX_STRENGTH = -2., 2.
+ERRORS = frozenset(('invalid_loras', 'invalid_lora_strength', 'duplicate_lora',
+                    'too_many_loras', 'lora_not_found', 'invalid_lora_model'))
+MESSAGE_ZH = {
+    'invalid_loras': 'LoRA 设置有误，请重新选择文件。',
+    'invalid_lora_strength': 'LoRA 强度必须在 -2 到 2 之间。',
+    'duplicate_lora': '同一个 LoRA 只能选择一次，请移除重复项。',
+    'too_many_loras': '一张图片最多使用 8 个 LoRA，请减少数量。',
+    'lora_not_found': 'LoRA 文件找不到，请重新选择。',
+    'invalid_lora_model': '此 LoRA 不适用于图片生成，请换一个。',
+}
+MESSAGE_EN = {
+    'invalid_loras': 'The LoRA selection is invalid. Choose the files again.',
+    'invalid_lora_strength': 'LoRA strength must be between -2 and 2.',
+    'duplicate_lora': 'Select each LoRA only once. Remove the duplicate.',
+    'too_many_loras': 'Use at most eight LoRAs per picture.',
+    'lora_not_found': 'The LoRA file could not be found. Choose it again.',
+    'invalid_lora_model': 'This LoRA cannot be used for pictures. Choose another one.',
+}
+
+
+def validate_selection(value):
+    """Canonical active items, with the same checks for HTTP and in-process jobs."""
+    if not isinstance(value, list):
+        raise ValueError('invalid_loras')
+    if len(value) > MAX_LORAS:
+        raise ValueError('too_many_loras')
+    result, seen = [], set()
+    for row in value:
+        if not isinstance(row, dict) or set(row) - {'name', 'strength', 'enabled'}:
+            raise ValueError('invalid_loras')
+        name, strength = row.get('name'), row.get('strength', 1.)
+        if type(row.get('enabled', True)) is not bool:
+            raise ValueError('invalid_loras')
+        if (not isinstance(name, str) or not 1 <= len(name) <= MAX_NAME
+                or any(ord(c) < 32 for c in name) or ':' in name
+                or any(part in ('', '.', '..') for part in name.replace('\\', '/').split('/'))
+                or not name.lower().endswith('.safetensors')):
+            raise ValueError('invalid_loras')
+        if (type(strength) not in (int, float) or not MIN_STRENGTH <= strength <= MAX_STRENGTH
+                or not math.isfinite(strength)):
+            raise ValueError('invalid_lora_strength')
+        if not row.get('enabled', True):
+            continue
+        if name in seen:
+            raise ValueError('duplicate_lora')
+        seen.add(name)
+        # Match resolve_loras: zero strength does not load weights or add triggers.
+        if strength != 0:
+            result.append(dict(name=name, strength=float(strength)))
+    return result
+
+
+def check_file(path):
+    """Image counterpart of lora_check.check_installed_file, without exposing paths."""
+    from .qwen_lora import inspect_file
+    try:
+        result = inspect_file(path)
+    except (OSError, ValueError):
+        result = dict(ok=False, reason='unreadable')
+    usable = bool(result['ok'])
+    return dict(usable=usable, reason=None if usable else result['reason'], precision=None,
+                **({} if usable else {'message_zh': '此 LoRA 不适用于图片生成。'}))
+
+
+def resolve_paths(value, *, folders=None):
+    rows = validate_selection(value)
+    if not rows:
+        return []
+    if folders is None:
+        import folder_paths as folders
+    names = set(folders.get_filename_list('loras'))
+    roots = [Path(root).resolve() for root in folders.get_folder_paths('loras')]
+
+    def resolve(name):
+        try:
+            path = Path(folders.get_full_path_or_raise('loras', name)).resolve(strict=True)
+        except (OSError, ValueError, KeyError):
+            raise ValueError('lora_not_found') from None
+        if not any(path.is_relative_to(root) for root in roots):
+            raise ValueError('invalid_loras')
+        if not path.is_file():
+            raise ValueError('lora_not_found')
+        if path.suffix.lower() != '.safetensors':
+            raise ValueError('invalid_loras')
+        return path
+
+    if any(row['name'] not in names for row in rows):
+        raise ValueError('lora_not_found')
+    from ..comfy_assets import resolve_loras
+    resolved = resolve_loras(json.dumps(rows), names, resolve)
+    return [dict(name=row['name'], **item) for row, item in zip(rows, resolved)]
+
+
+def resolve_selection(value, *, folders=None):
+    result = resolve_paths(value, folders=folders)
+    for item in result:
+        if not check_file(item['path'])['usable']:
+            raise ValueError('invalid_lora_model')
+    return result
+
+
+def check_named_file(name, *, folders=None):
+    return check_file(resolve_paths([dict(name=name)], folders=folders)[0]['path'])
diff --git a/freevideo_engine/image_gen/qwen_lora.py b/freevideo_engine/image_gen/qwen_lora.py
new file mode 100644
index 00000000..38f544cb
--- /dev/null
+++ b/freevideo_engine/image_gen/qwen_lora.py
@@ -0,0 +1,208 @@
+"""Community LoRAs for the Qwen-Image 2.1 DiT, applied as unmerged low-rank branches.
+
+The DiT's Linears are int8 ConvRot matrices, which cannot absorb a weight delta,
+so every LoRA stays a pair of small matrices and adds ``scale * B(A x)`` to the
+projection's output (the same shape as the video model's online LoRA). Nothing
+is written into the model's weights: attaching adds a forward wrapper per
+target, detaching restores the original forward, and a request without LoRAs
+runs the untouched model.
+
+Accepted files are the layouts the community trains for this model: kohya
+(``lora_down``/``lora_up`` and ``alpha``) and PEFT / ai-toolkit (``lora_A`` /
+``lora_B``), under a ``diffusion_model.``, ``transformer.`` or
+``base_model.model.`` prefix, or kohya's flat ``lora_unet_...`` names. The fused
+``img_mlp.gate_up`` target is split into the model's ``gate_layer`` (first half
+of the rows) and ``proj`` (second half), the order the checkpoint loader uses.
+"""
+import types
+from functools import lru_cache
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+
+UNUSABLE = 'This LoRA was not trained for the Qwen image model and cannot be used.'
+
+_PREFIXES = ('model.diffusion_model.', 'diffusion_model.', 'base_model.model.', 'transformer.')
+_DOWN = ('.lora_down.weight', '.lora_A.weight', '.lora_A.default.weight', '.lora_down.default.weight')
+_UP = ('.lora_up.weight', '.lora_B.weight', '.lora_B.default.weight', '.lora_up.default.weight')
+_ALPHA = ('.alpha', '.lora_alpha')
+_CHUNK = 512
+
+
+@lru_cache(maxsize=1)
+def layout():
+    """name -> (in_features, out_features) of every LoRA target; ``gate_up`` is the fused pair, added beside its two halves."""
+    from .qwen_dit import QwenImageDiT
+    from .qwen_weights import _config
+    with torch.device('meta'):
+        model = QwenImageDiT(**_config({}, QwenImageDiT, None))
+    shapes = {}
+    for name, module in model.named_modules():
+        if isinstance(module, torch.nn.Linear):
+            shapes[name] = (module.in_features, module.out_features)
+    for name in [n for n in shapes if n.endswith('.img_mlp.gate_layer')]:
+        prefix = name[:-len('gate_layer')]
+        inputs, rows = shapes[name]
+        shapes[prefix + 'gate_up'] = (inputs, rows * 2)
+    return shapes
+
+
+@lru_cache(maxsize=1)
+def _flat_names():
+    return {'lora_unet_' + name.replace('.', '_'): name for name in layout()}
+
+
+def _split(key):
+    """-> (target name, 'down' | 'up' | 'alpha') or None for a key that is not a LoRA tensor."""
+    for kind, suffixes in (('down', _DOWN), ('up', _UP), ('alpha', _ALPHA)):
+        for suffix in suffixes:
+            if key.endswith(suffix):
+                base = key[:-len(suffix)]
+                for prefix in _PREFIXES:
+                    if base.startswith(prefix):
+                        base = base[len(prefix):]
+                        break
+                return _flat_names().get(base, base), kind
+    return None
+
+
+def _entries(keys):
+    """Group tensor keys by target: name -> {'down': key, 'up': key, 'alpha': key}."""
+    groups = {}
+    for key in keys:
+        parsed = _split(key)
+        if parsed is not None:
+            groups.setdefault(parsed[0], {})[parsed[1]] = key
+    return groups
+
+
+def inspect_file(path):
+    """Header-only check: ``dict(ok, reason)``; reason is 'other-model', 'shape', 'empty' or 'unreadable'."""
+    from safetensors import safe_open
+    try:
+        with safe_open(str(path), framework='pt', device='cpu') as handle:
+            keys = list(handle.keys())
+            groups = _entries(keys)
+            if not keys:
+                return dict(ok=False, reason='empty')
+            shapes = layout()
+            mapped = {name: group for name, group in groups.items() if name in shapes}
+            if not mapped or not all('down' in g and 'up' in g for g in mapped.values()):
+                return dict(ok=False, reason='other-model')
+            for name, group in mapped.items():
+                inputs, outputs = shapes[name]
+                if 'alpha' in group:
+                    alpha = handle.get_tensor(group['alpha']).float()
+                    if alpha.numel() != 1 or not torch.isfinite(alpha).all():
+                        return dict(ok=False, reason='shape')
+                down = tuple(handle.get_slice(group['down']).get_shape())
+                up = tuple(handle.get_slice(group['up']).get_shape())
+                if len(down) != 2 or len(up) != 2 or down[1] != inputs or up[0] != outputs or down[0] != up[1] or down[0] < 1:
+                    return dict(ok=False, reason='shape')
+            return dict(ok=True, reason=None)
+    except Exception:
+        return dict(ok=False, reason='unreadable')
+
+
+def read(path, strength=1.0):
+    """-> list of (target, a [r, in], b [out, r], scale) in bf16 on the CPU. ValueError when unusable."""
+    from safetensors import safe_open
+    verdict = inspect_file(path)
+    if not verdict['ok']:
+        raise ValueError(UNUSABLE)
+    shapes = layout()
+    result = []
+    with safe_open(str(path), framework='pt', device='cpu') as handle:
+        for name, group in sorted(_entries(handle.keys()).items()):
+            if name not in shapes:
+                continue   # norm deltas and the like: not part of the supported targets
+            a = handle.get_tensor(group['down']).to(torch.bfloat16).contiguous()
+            b = handle.get_tensor(group['up']).to(torch.bfloat16).contiguous()
+            rank = a.shape[0]
+            alpha = float(handle.get_tensor(group['alpha']).float()) if 'alpha' in group else float(rank)
+            result.append((name, a, b, float(strength) * alpha / rank))
+    return result
+
+
+def file_bytes(loras):
+    """Bytes the branches of these ``{'path', 'strength'}`` entries occupy once read (bf16)."""
+    total = 0
+    for item in loras:
+        total += Path(item['path']).stat().st_size
+    return total
+
+
+class Branch:
+    """One LoRA's factors for one target; ``a``/``b`` live on ``home`` and move to the activation on use."""
+
+    __slots__ = ('a', 'b', 'scale')
+
+    def __init__(self, a, b, scale):
+        self.a, self.b, self.scale = a, b, scale
+
+
+def _apply(module, raw, output):
+    rows = raw.reshape(-1, raw.shape[-1])
+    target = output.reshape(-1, output.shape[-1])
+    if not target.is_contiguous():
+        raise ValueError('LoRA projection output must be contiguous')
+    for branch in module._qwen_lora:
+        a, b = branch.a, branch.b
+        if a.device != rows.device:
+            a, b = a.to(rows.device, non_blocking=True), b.to(rows.device, non_blocking=True)
+        for start in range(0, len(rows), _CHUNK):
+            z = F.linear(rows[start:start + _CHUNK].to(a.dtype), a)
+            target[start:start + _CHUNK].add_(F.linear(z, b).to(target.dtype), alpha=branch.scale)
+    return output
+
+
+class LoraSet:
+    """The branches of one request's LoRAs, attached to a DiT for the length of a with-block."""
+
+    def __init__(self, loras, *, device='cpu', pin=True):
+        self.entries = {}     # module name -> [Branch]
+        self.bytes = 0
+        self._wrapped = []
+        for item in loras:
+            for name, a, b, scale in read(item['path'], item['strength']):
+                if device != 'cpu':
+                    a, b = a.to(device), b.to(device)
+                elif pin and torch.cuda.is_available():
+                    try:
+                        a, b = a.pin_memory(), b.pin_memory()
+                    except RuntimeError:
+                        pass
+                self.bytes += a.numel() * a.element_size() + b.numel() * b.element_size()
+                if name.endswith('.img_mlp.gate_up'):
+                    prefix = name[:-len('gate_up')]
+                    half = b.shape[0] // 2
+                    self.entries.setdefault(prefix + 'gate_layer', []).append(Branch(a, b[:half], scale))
+                    self.entries.setdefault(prefix + 'proj', []).append(Branch(a, b[half:], scale))
+                else:
+                    self.entries.setdefault(name, []).append(Branch(a, b, scale))
+
+    def attach(self, model):
+        for name, branches in self.entries.items():
+            module = model.get_submodule(name)
+            module._qwen_lora = branches
+            original = module.forward
+
+            def forward(self, value, old=original):
+                return _apply(self, value, old(value))
+            module.forward = types.MethodType(forward, module)
+            self._wrapped.append(module)
+        return self
+
+    def detach(self):
+        for module in self._wrapped:
+            module.__dict__.pop('forward', None)
+            module.__dict__.pop('_qwen_lora', None)
+        self._wrapped = []
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        self.detach()
+        return False
diff --git a/freevideo_engine/image_gen/qwen_pipeline.py b/freevideo_engine/image_gen/qwen_pipeline.py
index 475cfa6b..e724fe08 100644
--- a/freevideo_engine/image_gen/qwen_pipeline.py
+++ b/freevideo_engine/image_gen/qwen_pipeline.py
@@ -44,6 +44,9 @@
 from .qwen_text import QwenTextEncoder
 
 
+# Free card memory that must remain after a request's LoRA branches are placed on the card; below it they
+# stay in pinned host memory and move to the card per use.
+LORA_GPU_MARGIN = 2 << 30
 TURBO_SIGMAS = (1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0)
 
 
@@ -257,7 +260,7 @@ def _reference_inputs(self, images):
 
     @torch.no_grad()
     def generate(self, prompt: str, *, width: int, height: int, seed: int | None = None, steps: int = 8,
-                 reference_images=None, edit_mask=None, progress=None, cancel=None) -> Image.Image:
+                 reference_images=None, edit_mask=None, loras=None, progress=None, cancel=None) -> Image.Image:
         if self._closed:
             raise RuntimeError("Generator is closed")
         if not isinstance(prompt, str):
@@ -269,6 +272,10 @@ def generate(self, prompt: str, *, width: int, height: int, seed: int | None = N
         width, height = normalize_dimensions(width, height)
         if edit_mask is not None and not reference_images:
             raise ValueError("edit_mask needs a reference image: the first one is the picture being edited")
+        if loras:
+            from .qwen_lora import UNUSABLE, inspect_file
+            if not all(inspect_file(item["path"])["ok"] for item in loras):
+                raise ValueError(UNUSABLE)
 
         def check_cancel():
             if cancel is not None and cancel():
@@ -347,7 +354,7 @@ def report(stage, fraction):
         image_mask = torch.cat([image_mask, image_mask.new_ones(1, latents.shape[1] // 4)], dim=1)
         sigmas = torch.tensor(self.sample_sigmas, device=self.device, dtype=torch.float32)
         timesteps = sigmas[:-1] * 1000
-        with self._stage("denoise"):
+        with self._stage("denoise"), self._lora_scope(loras):
             report("denoise", 0.0)
             for i, timestep in enumerate(timesteps):
                 check_cancel()
@@ -395,6 +402,23 @@ def report(stage, fraction):
             report("decode", 1.0)
         return image
 
+    @contextmanager
+    def _lora_scope(self, loras):
+        """Attach this request's LoRAs to the DiT for the denoising loop only; no LoRAs, no change."""
+        if not loras:
+            yield
+            return
+        from .qwen_lora import LoraSet
+        on_card = False
+        if torch.device(self.device).type == "cuda":
+            floor = LORA_GPU_MARGIN + sum(Path(item["path"]).stat().st_size for item in loras)
+            on_card = torch.cuda.mem_get_info(self.device)[0] > floor
+        branches = LoraSet(loras, device=self.device if on_card else "cpu")
+        try:
+            yield branches.attach(self.transformer)
+        finally:
+            branches.detach()
+
     def close(self) -> None:
         """Release owned model references; all active stages have already left."""
         self.text_encoder = self.transformer = self.vae = None
diff --git a/freevideo_engine/image_gen/service.py b/freevideo_engine/image_gen/service.py
index 3255b9fc..8889c7e6 100644
--- a/freevideo_engine/image_gen/service.py
+++ b/freevideo_engine/image_gen/service.py
@@ -16,7 +16,8 @@
 import time
 import uuid
 
-from . import image_pack
+from . import image_pack, keep
+from .loras import ERRORS as LORA_ERRORS, resolve_selection
 from ..download_settings import observed_speed
 from .failure import (Context, HOST_MESSAGE_ZH, host_memory_shortage, ERRORS, FailureError, MESSAGE_ZH, NO_REPORT, STAGES, UNIFIED_MESSAGE_ZH,
                       classify_exception)
@@ -34,7 +35,7 @@
 QUEUE_TIMEOUT = 600.
 QUEUE_POLL = .1
 INVALID_ERRORS = frozenset(('invalid_request', 'invalid_prompt', 'invalid_dimensions', 'invalid_seed',
-                            'invalid_media', 'invalid_points', 'too_many_images'))
+                            'invalid_media', 'invalid_points', 'too_many_images')) | LORA_ERRORS
 
 
 def comfy_server():
@@ -112,9 +113,10 @@ def close(self):
 
 
 class Service:
-    def __init__(self, installation, input_directory, *, video_busy=None, idle_seconds=45.):
+    def __init__(self, installation, input_directory, *, video_busy=None, idle_seconds=45., folders=None):
         self.installation = installation
         self.input_directory = input_directory
+        self.folders = folders
         self.video_busy = video_busy or self._video_busy
         self.idle_seconds = idle_seconds
         self.mutex = threading.RLock()
@@ -135,6 +137,7 @@ def __init__(self, installation, input_directory, *, video_busy=None, idle_secon
         self.worker_restarts = 0
         self.worker_started = False
         self.queued = None
+        self.habit_root = None
         self.queue_thread = None
         self.queue_wake = threading.Event()
         self.finished = OrderedDict()
@@ -501,9 +504,18 @@ def request(self, value):
         cast_view = value.get('cast_view')
         if cast_view is not None and (cast_view != 'back' or len(paths) != 1 or mask_path is not None):
             raise ValueError('invalid_request')
-        return dict(prompt=prompt, width=dimensions[0] // 32 * 32, height=dimensions[1] // 32 * 32,
-                    steps=8, seed=seed, images=paths, **(dict(mask=mask_path) if mask_path else {}),
-                    **(dict(cast_view='back') if cast_view == 'back' else {}), path=self.output_path(named=True))
+        loras = resolve_selection(value.get('loras', []), folders=self.folders)
+        if loras:
+            from ..lora_triggers import apply_triggers
+            prompt = apply_triggers(prompt, loras, folders=self.folders)
+        result = dict(prompt=prompt, width=dimensions[0] // 32 * 32, height=dimensions[1] // 32 * 32,
+                      steps=8, seed=seed, images=paths, **(dict(mask=mask_path) if mask_path else {}),
+                      **(dict(cast_view='back') if cast_view == 'back' else {}), path=self.output_path(named=True),
+                      **(dict(loras=loras) if loras else {}))
+        # Leave room for action, kind and the small size-fit record added at dispatch.
+        if loras and len(json.dumps(result, ensure_ascii=True).encode('utf-8')) > BODY_LIMIT - 1024:
+            raise ValueError('invalid_loras')
+        return result
 
     def start(self, action, value):
         if action not in ('generate', 'compose'):
@@ -882,6 +894,7 @@ def spawn(self, root, machine, env, descriptor):
     def work(self, root, machine):
         from ..locking import runtime_lock
         from .. import processes, resident_process
+        self.habit_root = root
         session = None
         failure = 'worker_exit'
         failure_error = None
@@ -1082,6 +1095,11 @@ def worker_gone_error(session, heard):
     def infer(self, session, value):
         context = self.context or Context('generate', value)
         heard = False
+        habit = keep.path(self.habit_root)
+        if 'keep_seconds' not in value:
+            keep.note('image', habit)
+        # How long the worker may wait for the next picture; it applies that only with the whole model on the card.
+        value['keep_seconds'] = keep.keep_seconds(habit)
         try:
             session.send(action='generate', **value)
         except OSError:
@@ -1194,7 +1212,7 @@ def video_busy():
         running, pending = server.prompt_queue.get_current_queue()
         return bool(running or pending) or Service._video_busy()
 
-    service = Service(installation, folder_paths.get_input_directory, video_busy=video_busy)
+    service = Service(installation, folder_paths.get_input_directory, video_busy=video_busy, folders=folder_paths)
     server._freevideo_image_gen = service
     token = secrets.token_urlsafe(24)
 
diff --git a/freevideo_engine/image_gen/worker.py b/freevideo_engine/image_gen/worker.py
index f9c565a3..33bfc283 100644
--- a/freevideo_engine/image_gen/worker.py
+++ b/freevideo_engine/image_gen/worker.py
@@ -9,6 +9,7 @@
 import threading
 import time
 
+from . import keep
 from .failure import Context, classify_exception, exception_details, host_memory_shortage, safe_plan
 
 
@@ -71,6 +72,35 @@ def _flatten(image):
     return image.convert('RGB')
 
 
+def next_request(inbox, wait, crowding=None):
+    """The next row, or None when the pipe closed, `wait` seconds passed, or the card is crowded.
+
+    Without `crowding` this is one blocking read, as for an ordinary idle wait."""
+    if crowding is None:
+        try:
+            return inbox.rows.get(timeout=wait)
+        except queue.Empty:
+            return None
+    deadline = time.monotonic() + wait
+    while True:
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            return None
+        try:
+            return inbox.rows.get(timeout=min(remaining, keep.POLL_SECONDS))
+        except queue.Empty:
+            if crowding:
+                return None
+
+
+def hold(runner):
+    """Start the long idle wait: hand back the allocator's cached memory, then watch the card."""
+    torch = sys.modules.get('torch')
+    if torch is not None:
+        torch.cuda.empty_cache()
+    return keep.Crowding(lambda: keep.card_free_bytes(runner.device))
+
+
 def load_runner(pack_dir, device, *, progress=None):
     # Importing this module in ComfyUI or in CPU protocol tests never loads torch.
     from .image_runner import ImageRunner
@@ -92,12 +122,10 @@ def send(**row):
     runner = None
     back_detector = None
     back_detector_loaded = False
+    wait, crowding = idle_seconds, None
     try:
         while not inbox.closed.is_set():
-            try:
-                value = inbox.rows.get(timeout=idle_seconds)
-            except queue.Empty:
-                break
+            value = next_request(inbox, wait, crowding)
             if value is None or inbox.closed.is_set():
                 break
             started = time.monotonic()
@@ -126,11 +154,13 @@ def progress(stage, fraction):
                     if value.get('mask'):
                         region['edit_mask'] = stack.enter_context(Image.open(value['mask']))
                         region['edit_mask'].load()
+                    loras = value.get('loras') or None
                     back_check = None
                     def generate(seed):
                         return runner.generate(
                             value['prompt'], width=value['width'], height=value['height'],
                             seed=seed, reference_images=references, **region,
+                            **(dict(loras=loras) if loras else {}),
                             progress=progress, cancel=inbox.cancel.is_set)
 
                     if value.get('cast_view') == 'back' and len(references) == 1 and not region:
@@ -195,6 +225,16 @@ def generate(seed):
                     inbox.busy = False
                     if result is not None:
                         send(**result)
+            wait, crowding = idle_seconds, None
+            if result is not None and result.get('phase') == 'result' and runner is not None:
+                wait = keep.wait_seconds(value.get('keep_seconds'), runner, idle_seconds)
+                if wait > idle_seconds:
+                    try:
+                        crowding = hold(runner)
+                    except Exception as error:
+                        # The picture is already sent; a hold that cannot start is the normal short wait.
+                        print('image worker hold failed: %s' % type(error).__name__, file=sys.stderr, flush=True)
+                        wait = idle_seconds
     finally:
         inbox.closed.set()
         inbox.cancel.set()
diff --git a/freevideo_engine/media_encoding.py b/freevideo_engine/media_encoding.py
index c2da401e..c3a7a446 100644
--- a/freevideo_engine/media_encoding.py
+++ b/freevideo_engine/media_encoding.py
@@ -351,9 +351,12 @@ def encode_latents(normalized, value, base, canvas, *, resident=None, device='cu
         on_mps = torch.device(device).type == 'mps'
         # On MPS the weights stay in the checkpoint mapping and each encoder
         # stage uploads its own while it runs (bounded_encoder).
-        vae, _ = load_video_encoder(base, before_upload=None if resident is None else
-                                    lambda model: resident.input_vae_room(model, canvas),
-                                    device='cpu' if on_mps else device)
+        if resident is not None and not on_mps:
+            vae = resident.input_video_encoder(base, canvas, device=device)
+        else:
+            vae, _ = load_video_encoder(base, before_upload=None if resident is None else
+                                        lambda model: resident.input_vae_room(model, canvas),
+                                        device='cpu' if on_mps else device)
         lifetime = nullcontext()
         if on_mps:
             from .backends.mps_vae_encode import bounded_encoder
@@ -382,8 +385,9 @@ def encode_latents(normalized, value, base, canvas, *, resident=None, device='cu
         empty_cache()
     if any(row.get('audio') for row in normalized):
         print(json.dumps({'event': 'media_encode_phase', 'phase': 'Encoding reference audio'}), flush=True)
-        vae, _ = load_audio_vae(base, before_upload=None if resident is None else
-                                lambda model: resident.input_vae_room(model, canvas, audio=True), device=device)
+        vae = (resident.input_audio_encoder(base, canvas, device=device)
+               if resident is not None and torch.device(device).type != 'mps' else
+               load_audio_vae(base, device=device)[0])
         mean = torch.tensor(vae.config.latents_mean).view(1, 1, -1)
         std = torch.tensor(vae.config.latents_std).view(1, 1, -1)
         if vae.config.sampling_rate != 32000:
diff --git a/freevideo_engine/release_notes.json b/freevideo_engine/release_notes.json
index 64913c72..cb8bc022 100644
--- a/freevideo_engine/release_notes.json
+++ b/freevideo_engine/release_notes.json
@@ -1,41 +1,68 @@
 {
   "schema": 1,
-  "product_version": "0.4.27",
+  "product_version": "0.4.28",
   "en": {
-    "summary": "The Agent can make music for a project, an update started during a generation waits for it to finish, and a preview can keep the guide clip's sound.",
+    "summary": "Image generation, editing and character generation accept LoRAs, the image model stays on the card longer for people who mostly make pictures, and Repair replaces Python packages whose files are missing.",
     "changes": [
-      "Agent: a new music tool makes a soundtrack for a project and places it on an audio track. The music model is downloaded on first use (about 20 GB); cards with little video memory use a built-in 8-bit form. If two attempts fail, the Agent synthesizes the music itself.",
-      "Updates: when \"Update now\" is pressed while a video is generating, the page shows a waiting screen instead of a failure card, and the update starts when the job ends. The wait also covers a job on a server that another program started.",
-      "Audio: a kept original audio track is written as lossless audio that browsers can play. A preview made with a guide clip keeps the clip's sound when that was requested.",
-      "Long videos: when seams cannot be joined, the notes name the seams that failed first, with the step and the format. The how-to for a long video points to \"Extend\" below the finished video.",
-      "Phone access: the password page is shown in Chinese when the browser's preferred language is Chinese.",
-      "Speed: the encoded text of a prompt is kept per encoder file and prompt, so a repeated prompt starts sooner.",
-      "Precompile: the status keeps the order of its writes, stopping responds quickly, and the diagnostic report includes the status.",
-      "Decoding and reports: a CUDA error while decoding after sampling has finished no longer fails the video; the saved result is decoded once more in a fresh process. A process that keeps running after a CUDA error is ended, so the failure is reported instead of waiting. Diagnostic reports record the polling of Agent generations.",
-      "Errors: a refused offline package says why, and a slow ComfyUI page is no longer blamed on the browser.",
-      "Install: on Windows the portable Git is used when the installed Git cannot load libcurl, and the Windows certificate store is used for HTTPS checks, skipping a certificate that cannot be read. A rejected token or address stays explained under its field until it is accepted.",
-      "Linux: on Arch, the setup names the pacman command for missing basic tools.",
-      "Mac: after two Metal command buffer failures during sampling, a third attempt runs with Torch attention."
+      "Image generation: image generation, image editing and character generation each have a LoRA control. Up to 8 LoRAs can be added, each with a file, a strength and a trigger word that is added to the prompt. Video and image LoRAs are told apart and a wrong choice is explained in one sentence. Without a LoRA the output is unchanged.",
+      "Speed: for people who mostly make pictures, the image model stays loaded for up to 10 minutes when it fits on the card, so the next picture starts sooner. It gives the card back at once when a video or another program needs it.",
+      "Speed: the usage habit that decides which model parts stay in memory (only the kind of each operation and its count, nothing from prompts or images) is now kept across restarts. When prompts change often, the text encoder is loaded while the engine is idle, as far as free video and system memory allow, and stops when a new job starts.",
+      "Speed: after the engine starts, the model files for the first request are read ahead as far as free memory allows: the video model for people who mostly generate again with the same prompt and images, and the text encoder otherwise. A new request or a busy queue stops the reading, and a machine with little free memory reads less.",
+      "Install: Repair now reinstalls Python packages whose files are missing or cut short, for example a numpy library that has gone missing, which setup did not notice before when the package was already at the required version; the repair review counts them in the download. When the hardware check before a request fails to load such a package, the message names the missing file and offers Repair.",
+      "Windows: on machines with little memory, a long video whose second pass stops for lack of system memory now retries in a fresh process, keeping the finished first pass, and can finish instead of failing.",
+      "Guide: the tutorial book covers the additions of 0.4.25 and 0.4.26 and has a new chapter on replacing a person, with a matching guide card. In the Chinese book, quoted interface names of up to eight characters and numbers with their units are no longer split across lines."
     ]
   },
   "zh": {
-    "summary": "Agent 可以为作品制作配乐；生成中点击更新会等任务结束后再更新；预览也可以保留引导片段的声音。",
+    "summary": "图片生成、改图和角色生成可以使用 LoRA；以做图为主时，图片模型在显存中保留更久；“修复”会重新安装文件缺失的 Python 包。",
     "changes": [
-      "Agent：新增配乐工具，为作品制作配乐并放到一条音频轨上。配乐模型首次使用时下载（约 20 GB）；显存较小的显卡使用内置的 8 位形式。两次尝试都失败时，由 Agent 自己合成。",
-      "更新：有视频正在生成时点击“立即更新”，页面显示等待画面，不再出现失败卡片，任务结束后开始更新。由其他程序启动的服务上的任务也会等待。",
-      "音频：保留的原始音轨以浏览器可播放的无损音频写入。使用引导片段的预览在要求保留时，也保留该片段的声音。",
-      "长视频：接缝无法衔接时，说明先列出失败的接缝及其步骤和格式。长视频的使用指引改为指向成片下方的“接着生成”。",
-      "手机访问：浏览器的首选语言为中文时，密码页显示中文。",
-      "速度：提示词的文本编码结果按编码器文件和提示词保存，重复使用同一提示词时更快开始。",
-      "预编译：状态按写入顺序保存，停止时响应更快，诊断报告包含该状态。",
-      "解码与诊断报告：采样完成后解码时出现 CUDA 错误，不再整条失败，已保存的结果会在新进程中重新解码一次；出现 CUDA 错误后仍不退出的进程会被结束，直接报告失败，不再一直等待。诊断报告记录 Agent 生成任务的轮询情况。",
-      "错误提示：离线包被拒绝时说明原因；ComfyUI 页面加载较慢时不再归咎于浏览器。",
-      "安装：Windows 上已安装的 Git 无法载入 libcurl 时改用便携版 Git；HTTPS 检查使用 Windows 证书库，遇到无法读取的证书时跳过。令牌或地址被拒绝后，说明一直显示在对应输入框下，直到输入被接受。",
-      "Linux：在 Arch 上缺少基础工具时，安装脚本给出对应的 pacman 命令。",
-      "Mac：采样时连续两次 Metal 命令缓冲区失败后，用 Torch 注意力再试第三次。"
+      "图片生成：图片生成、改图和角色生成各增加了 LoRA 控件。最多可添加 8 个 LoRA，每个可选择文件、调整强度，触发词会加入提示词。视频和图片的 LoRA 分开识别，选错时用一句话说明。不使用 LoRA 时输出不变。",
+      "速度：以做图为主的用户，图片模型在能完整放进显存时最多保留 10 分钟，下一张图开始更快；视频或其他程序需要显存时立即让出。",
+      "速度：决定哪些模型部分驻留内存的操作习惯（只记录每次操作的类型和次数，不含提示词和图片的任何内容）现在会跨重启保存。经常更换提示词时，引擎空闲期间会预先载入文本编码器，只载入空闲显存和内存放得下的部分，新任务开始即停止。",
+      "速度：引擎启动后，按空闲内存的允许范围预先读入第一次请求要用的模型文件：以相同提示词和图片反复生成为主时读视频模型，其余情况读文本编码器。新请求到来或队列忙碌时停止读取，空闲内存较少的机器读得更少。",
+      "安装：“修复”现在会重新安装文件缺失或被截短的 Python 包（例如 numpy 的某个库文件丢失）；此前包已是所需版本时，这类问题安装流程检测不到。修复审阅页会把它们计入下载量。请求前的硬件检测因此类包无法加载而失败时，失败信息会点名缺失的文件并提供“修复”。",
+      "Windows：内存较小的机器生成长视频时，第二遍因系统内存不足中断后，会保留已完成的第一遍，改在新进程里重试，可以继续完成，而不是整次失败。",
+      "教程：教程书补充了 0.4.25 和 0.4.26 新增功能的说明，并新增“替换人物”一章和对应的导览卡。中文教程书中，八个字以内的带引号界面名称以及数字与单位，不再被拆在两行。"
     ]
   },
   "history": [
+    {
+      "product_version": "0.4.27",
+      "en": {
+        "summary": "The Agent can make music for a project, an update started during a generation waits for it to finish, and a preview can keep the guide clip's sound.",
+        "changes": [
+          "Agent: a new music tool makes a soundtrack for a project and places it on an audio track. The music model is downloaded on first use (about 20 GB); cards with little video memory use a built-in 8-bit form. If two attempts fail, the Agent synthesizes the music itself.",
+          "Updates: when \"Update now\" is pressed while a video is generating, the page shows a waiting screen instead of a failure card, and the update starts when the job ends. The wait also covers a job on a server that another program started.",
+          "Audio: a kept original audio track is written as lossless audio that browsers can play. A preview made with a guide clip keeps the clip's sound when that was requested.",
+          "Long videos: when seams cannot be joined, the notes name the seams that failed first, with the step and the format. The how-to for a long video points to \"Extend\" below the finished video.",
+          "Phone access: the password page is shown in Chinese when the browser's preferred language is Chinese.",
+          "Speed: the encoded text of a prompt is kept per encoder file and prompt, so a repeated prompt starts sooner.",
+          "Precompile: the status keeps the order of its writes, stopping responds quickly, and the diagnostic report includes the status.",
+          "Decoding and reports: a CUDA error while decoding after sampling has finished no longer fails the video; the saved result is decoded once more in a fresh process. A process that keeps running after a CUDA error is ended, so the failure is reported instead of waiting. Diagnostic reports record the polling of Agent generations.",
+          "Errors: a refused offline package says why, and a slow ComfyUI page is no longer blamed on the browser.",
+          "Install: on Windows the portable Git is used when the installed Git cannot load libcurl, and the Windows certificate store is used for HTTPS checks, skipping a certificate that cannot be read. A rejected token or address stays explained under its field until it is accepted.",
+          "Linux: on Arch, the setup names the pacman command for missing basic tools.",
+          "Mac: after two Metal command buffer failures during sampling, a third attempt runs with Torch attention."
+        ]
+      },
+      "zh": {
+        "summary": "Agent 可以为作品制作配乐；生成中点击更新会等任务结束后再更新；预览也可以保留引导片段的声音。",
+        "changes": [
+          "Agent：新增配乐工具，为作品制作配乐并放到一条音频轨上。配乐模型首次使用时下载（约 20 GB）；显存较小的显卡使用内置的 8 位形式。两次尝试都失败时，由 Agent 自己合成。",
+          "更新：有视频正在生成时点击“立即更新”，页面显示等待画面，不再出现失败卡片，任务结束后开始更新。由其他程序启动的服务上的任务也会等待。",
+          "音频：保留的原始音轨以浏览器可播放的无损音频写入。使用引导片段的预览在要求保留时，也保留该片段的声音。",
+          "长视频：接缝无法衔接时，说明先列出失败的接缝及其步骤和格式。长视频的使用指引改为指向成片下方的“接着生成”。",
+          "手机访问：浏览器的首选语言为中文时，密码页显示中文。",
+          "速度：提示词的文本编码结果按编码器文件和提示词保存，重复使用同一提示词时更快开始。",
+          "预编译：状态按写入顺序保存，停止时响应更快，诊断报告包含该状态。",
+          "解码与诊断报告：采样完成后解码时出现 CUDA 错误，不再整条失败，已保存的结果会在新进程中重新解码一次；出现 CUDA 错误后仍不退出的进程会被结束，直接报告失败，不再一直等待。诊断报告记录 Agent 生成任务的轮询情况。",
+          "错误提示：离线包被拒绝时说明原因；ComfyUI 页面加载较慢时不再归咎于浏览器。",
+          "安装：Windows 上已安装的 Git 无法载入 libcurl 时改用便携版 Git；HTTPS 检查使用 Windows 证书库，遇到无法读取的证书时跳过。令牌或地址被拒绝后，说明一直显示在对应输入框下，直到输入被接受。",
+          "Linux：在 Arch 上缺少基础工具时，安装脚本给出对应的 pacman 命令。",
+          "Mac：采样时连续两次 Metal 命令缓冲区失败后，用 Torch 注意力再试第三次。"
+        ]
+      }
+    },
     {
       "product_version": "0.4.26",
       "en": {
@@ -197,39 +224,6 @@
           "页面检查：ComfyUI 的调暗或点击遮罩（首次启动的模板对话框、帮助面板等）盖在 FreeVideo 工具栏上时，不再报告工具栏被遮挡；诊断详情会写出遮挡工具栏的元素。"
         ]
       }
-    },
-    {
-      "product_version": "0.4.20",
-      "en": {
-        "summary": "One continuous shot can now be split into up to 8 parts; generated pictures can be downloaded directly and the path of a video can be copied.",
-        "changes": [
-          "Long videos: one continuous shot can now be split into up to 8 parts instead of 5. The line below the duration in the creation panel states that each run makes one continuous shot and points to Edit or Autopilot for a storyboard; a length that needs more parts than allowed is stopped before it is submitted, and a refusal at submission is shown as a plain notice instead of an unknown error.",
-          "Reference video: fixed generation failing when a reference video is combined with a non-default upscale factor such as 1.5x.",
-          "Pictures and videos: picture cards in the Autopilot chat and the enlarged picture view have a Download button; results in the library and the creation panel have Copy path.",
-          "The desktop notification for a finished generation is now on by default: permission is requested once, the first time Generate is clicked; it stays off for anyone who has turned it off.",
-          "Edit: while the generation service has not answered, the generation bar stays visible with Generate dimmed; when the automatic checks get no answer, the status line shows \"Generation unavailable\" with a retry link, which checks again when clicked.",
-          "The edits, imports, characters, Agent conversations and Agent jobs lists now end with a line such as \"N edits could not be read\" when some records cannot be read.",
-          "Windows graphics cards with 8 to 12 GB of video memory: the first pass of FP8 two-pass videos on RTX 40 and 50 series cards, and the second pass of int8 videos longer than 10 seconds on 12 GB cards, now run attention 8 heads at a time instead of 4, within the planned memory budget. LoRA, Linux and Mac are unchanged. The same seed can give a slightly different video.",
-          "Retry after one-time errors: a second pass that stops with a one-time CUDA error resumes once from the finished first pass instead of failing or repeating it; on a Mac, a pass that stops with a Metal command-buffer error is retried once with smaller memory partitions. Failure reports name the machine, settings and retry result, and report ids stay retrievable after the service restarts.",
-          "Decoding the same request with the same seed now writes identical pixels every time. Compared with earlier versions, the same seed can give slightly different pixels.",
-          "The card shown for an unsupported graphics card now names the detected card and its compute capability, and states that FreeVideo needs compute capability 8.0 or higher (GeForce RTX 30 series or newer) regardless of VRAM. Error text no longer replaces a slash inside a sentence with a path placeholder."
-        ]
-      },
-      "zh": {
-        "summary": "一个连续镜头最多可分 8 段；生成的图片可以直接下载，作品可以复制路径。",
-        "changes": [
-          "长视频：一个连续镜头最多可分 8 段（原为 5 段）。创作面板时长下方写明每次生成一个连续镜头，有分镜剧本时可转到剪辑视图或 AI 全自动；所需段数超过上限时在提交前提示缩短时长，提交时仍被拒绝也显示为普通提示，不再显示为未知错误。",
-          "参考视频：修复参考视频配合 1.5 倍等非默认放大倍数时生成失败。",
-          "图片和作品：AI 全自动对话中的图片卡和放大查看窗增加“下载”；作品库和创作面板的结果按钮增加“复制路径”。",
-          "生成完成的桌面通知改为默认开启：第一次点生成时请求一次通知权限；已经关闭通知的保持关闭。",
-          "剪辑视图：生成服务暂时没有回答时，生成栏不再被藏起，“生成”按钮置灰；几次自动检查都没有回答时，状态行显示“生成服务暂不可用 · 重试”，点击后重新检查。",
-          "剪辑、已导入、角色、Agent 对话和 Agent 任务列表中有记录无法读取时，列表末尾写明“另有 N 个……无法读取”。",
-          "Windows 上显存 8–12 GB 的显卡：RTX 40 和 50 系列 FP8 两遍视频的第一遍，以及 12 GB 显卡上超过 10 秒的 int8 视频的第二遍，每次计算的注意力头由 4 个增加到 8 个，显存占用保持在计划预算内。LoRA、Linux 和 Mac 不变。同一个种子生成的视频可能略有不同。",
-          "一次性错误自动重试：第二遍遇到一次性的 CUDA 错误时，自动从已完成的第一遍续跑一次，不再整个请求失败或重做第一遍；Mac 上采样阶段遇到 Metal 命令缓冲区错误时，会用更小的内存分块重试一次。失败报告写明机器、设置和重试结果，报告编号在服务重启后仍可取回。",
-          "同一请求、同一种子，解码出的画面每次逐位一致。与之前的版本相比，同一种子的画面可能有细微差别。",
-          "不受支持的显卡的提示卡现在写出检测到的显卡和计算能力，并说明 FreeVideo 需要计算能力 8.0 及以上的显卡（GeForce RTX 30 系列及更新），与显存大小无关。错误文字中句子里的斜杠不再被替换成路径占位符。"
-        ]
-      }
     }
   ]
 }
diff --git a/freevideo_engine/resident_models.py b/freevideo_engine/resident_models.py
index b55c139c..7f03200a 100644
--- a/freevideo_engine/resident_models.py
+++ b/freevideo_engine/resident_models.py
@@ -1,17 +1,83 @@
 """Request-aware model residency inside the isolated interactive GPU worker."""
 import gc
+import hashlib
 import json
 from pathlib import Path
 import sys
 import time
 
 import torch
+from . import usage_habit
 from .torch_compat import empty_host_cache, host_cache_release_supported
 from .device_index import device_properties
 
 GiB = 1 << 30
 # Where a preview's sampling peak is kept, apart from its complete request's.
 PREVIEW_STAGE = 'engine_preview'
+HABIT_WINDOW = 4
+
+
+def retention_order(history):
+    """Cache roles to evict first; two thirds of recent transitions wins.
+
+    History contains at most four successful requests' digests, never inputs.
+    Image changes win ties with prompt changes, then same-prompt reruns. With
+    fewer than two successes (or no dominant habit), use the old engine-first
+    order. Active stages still use make_room's existing admission rules.
+    """
+    return eviction_order(usage_habit.dominant(usage_habit.transitions(history)))
+
+
+def eviction_order(habit):
+    """Eviction order for a habit kind from usage_habit.dominant (None: default)."""
+    if habit == 'images':
+        keep = ('input_video_vae', 'encoder', 'conditioning', 'engine',
+                'input_audio_vae', 'video_vae', 'audio_vae')
+    elif habit == 'prompt':
+        keep = ('encoder', 'conditioning', 'engine', 'input_video_vae',
+                'input_audio_vae', 'video_vae', 'audio_vae')
+    elif habit == 'rerun':
+        keep = ('conditioning', 'engine', 'video_vae', 'audio_vae',
+                'input_video_vae', 'input_audio_vae', 'encoder')
+    else:
+        return ('engine', 'encoder', 'audio_vae', 'video_vae',
+                'input_audio_vae', 'input_video_vae', 'conditioning')
+    return tuple(reversed(keep))
+
+
+def _input_digest(value):
+    """Hash CPU conditioning without retaining inputs or copying mapped payloads."""
+    digest = hashlib.sha256()
+    def visit(item):
+        if isinstance(item, torch.Tensor):
+            tensor = item.detach().cpu().contiguous()
+            digest.update(str((str(tensor.dtype), list(tensor.shape))).encode())
+            digest.update(memoryview(tensor.reshape(-1).view(torch.uint8).numpy()))
+        elif isinstance(item, dict):
+            for name in sorted(item):
+                visit(name)
+                visit(item[name])
+        elif isinstance(item, (tuple, list)):
+            for child in item:
+                visit(child)
+        else:
+            digest.update(json.dumps(item, sort_keys=True).encode())
+        digest.update(b'\0')
+    visit(value)
+    return digest.hexdigest()
+
+
+class ConditioningMemo(torch.nn.Module):
+    """The engine's bounded CPU refined-text results, independently evictable."""
+    def __init__(self, memo):
+        super().__init__()
+        self.memo = memo
+
+    def to(self, device):
+        if device != 'meta':
+            raise ValueError('Conditioning memo stays on the CPU')
+        self.memo.clear()
+        return self
 
 
 def key(value):
@@ -156,6 +222,11 @@ def __init__(self):
         from .streamed_weights import ConvertedStore
         # Converted decoder layers streamed from the checkpoint; rebuilt on demand.
         self.converted_store = ConvertedStore()
+        self.history = []
+        # Transition kinds kept across restarts (see usage_habit); history
+        # holds only this process's digests.
+        stored = usage_habit.load()
+        self.recent = stored['recent'] if stored else []
         # Set by a request that can read the whole device: bytes no process
         # holds, and the growth reserve that request keeps out of them. The
         # planning budget already excludes that reserve; the device reading
@@ -200,6 +271,11 @@ def drop(self, role, reason):
             import comfy.model_management
             encoder_discarded = release_encoder(model, comfy.model_management)
         elif role == 'engine':
+            cached = self.entries.get('conditioning')
+            if cached and getattr(model, 'refined_text_memo', None) is cached['model'].memo:
+                # Engine.close clears its dictionary; detach the independently
+                # owned results so an engine miss need only reload weights.
+                model.refined_text_memo = {}
             model.close()
         else:
             model.to('meta')
@@ -268,14 +344,15 @@ def ram_fits(self, need=2*GiB):
                 (self.ram_budget is None or used + need <= self.ram_budget))
 
     def settle(self, pinned=0):
-        """End of a request: keep the engine and encoder only if the next cold encoder still fits.
+        """Keep as many independent components as the next cold start permits.
 
         The RAM guard counts the whole session, so retained engine weights
         plus the next encoder's start (about 8 GiB) must stay inside the
         budget, and so must the host weights a fresh plan would pin beyond those
         the engine already holds: measured, a retained copy left no room to
-        pin and sampling took 95 s instead of 38 s. Otherwise the engine goes
-        as it did with one worker per request.
+        pin and sampling took 95 s instead of 38 s. Evict lower-priority
+        components first, checking live RAM again after each release. This does
+        not change an engine's layer layout.
         """
         # A retained engine's pinned weights are reused by the next request
         # rather than pinned again, so they count once, in the session's use.
@@ -283,8 +360,11 @@ def settle(self, pinned=0):
         need = 8*GiB + max(0, int(pinned) - held)
         if not self.ram_fits(need):
             self.converted_store.clear()
-        for role in ('engine', 'encoder'):
-            if role in self.entries and not self.ram_fits(need):
+        order = eviction_order(self.habit())
+        for role in (*order, *(role for role in self.entries if role not in order)):
+            if self.ram_fits(need):
+                break
+            if role in self.entries:
                 self.drop(role, 'Retained weights leave no host room for the next request')
                 if role == 'engine':
                     need = 8*GiB + int(pinned)
@@ -295,6 +375,37 @@ def settle(self, pinned=0):
         except (OSError, AttributeError):
             pass
 
+    def succeeded(self, request):
+        """Observe video successes only, locally; encoding/preloads are not jobs.
+
+        Read the existing conditioning with mmap after generation has returned,
+        so failures and cancellations do not train the policy. An unreadable
+        or legacy external input leaves the previous habit unchanged.
+        """
+        path = Path(request['conditioning'])
+        value = torch.load(path, map_location='cpu', weights_only=True, mmap=True)
+        prompt = value.get('prompt')
+        if prompt is None:
+            prompt = value['prompt_embeds'][value['text_token_tags'] == 1]
+        images = dict(anchors=value.get('keyframe_anchors'), latents=value.get('condition_latents'),
+                      references=[{k: row[k] for k in ('kind', 'latent') if k in row}
+                                  for row in value.get('references', []) if row['kind'] != 'audio'])
+        row = dict(prompt=_input_digest(prompt), images=_input_digest(images),
+                   run=key(dict(seed=request.get('seed'), options=request.get('engine_options'),
+                                geometry=request.get('geometry'), sampling=request.get('sampling_plan'))))
+        kinds = usage_habit.transition(self.history[-1], row) if self.history else None
+        self.history = (self.history + [row])[-HABIT_WINDOW:]
+        if kinds is not None:
+            self.recent = (self.recent + [kinds])[-(HABIT_WINDOW - 1):]
+        try:
+            usage_habit.record(kinds)
+        except OSError:
+            pass  # A habit that cannot be saved is only forgotten at restart.
+
+    def habit(self):
+        """The dominant kind of change, from stored transitions or this process's digests."""
+        return usage_habit.dominant(self.recent or usage_habit.transitions(self.history))
+
     def input_vae_room(self, model, canvas, *, audio=False):
         # Input encoding has its own live workspace; it cannot assume an output
         # decoder's memory footprint. Charge actual retained CPU model tensors
@@ -311,7 +422,41 @@ def input_vae_room(self, model, canvas, *, audio=False):
         # model's host copy is released unless the live budget covers the
         # stage as well. Plentiful hosts keep it for reuse.
         host = weights + (2*GiB if audio else max(8*GiB, 26*GiB*area))
-        self.make_room('input_audio_vae' if audio else 'input_video_vae', weights + workspace, ram_need=host)
+        role = 'input_audio_vae' if audio else 'input_video_vae'
+        retained = model if self.entries.get(role, {}).get('model') is model else None
+        self.make_room(role, weights + workspace, model=retained, ram_need=host)
+
+    def input_video_encoder(self, base, canvas, *, device='cuda'):
+        """Reuse just the image/video encoder; never load the output decoder."""
+        from .vae_weights import load_video_encoder
+        directory = Path(base) / 'vae'
+        identity = key(dict(files=[file_identity(path) for path in
+                           sorted([directory/'config.json', *directory.glob('*.safetensors')])],
+                           device=str(device)))
+        found = self.take('input_video_vae', identity)
+        if found is not None:
+            self.input_vae_room(found, canvas)
+            return found
+        found, _ = load_video_encoder(base, before_upload=lambda model: self.input_vae_room(model, canvas),
+                                     device=device)
+        self.put('input_video_vae', identity, found)
+        return found
+
+    def input_audio_encoder(self, base, canvas, *, device='cuda'):
+        """Reuse the reference-audio VAE the way the image/video encoder is reused."""
+        from .vae_weights import load_audio_vae
+        directory = Path(base) / 'audio_vae'
+        identity = key(dict(files=[file_identity(path) for path in
+                           sorted([directory/'config.json', *directory.glob('*.safetensors')])],
+                           device=str(device)))
+        found = self.take('input_audio_vae', identity)
+        if found is not None:
+            self.input_vae_room(found, canvas, audio=True)
+            return found
+        found, _ = load_audio_vae(base, before_upload=lambda model: self.input_vae_room(model, canvas, audio=True),
+                                  device=device)
+        self.put('input_audio_vae', identity, found)
+        return found
 
     def device_ceiling(self):
         """What this process's allocator can hold before the device itself is full.
@@ -450,6 +595,13 @@ def engine(self, request, factory):
         if found is None:
             found = factory()
             self.put('engine', identity, found)
+        memo = getattr(found, 'refined_text_memo', None)
+        if isinstance(memo, dict):
+            cached = self.take('conditioning', identity)
+            if cached is None:
+                cached = ConditioningMemo(memo)
+                self.put('conditioning', identity, cached)
+            found.refined_text_memo = cached.memo
         return found, hit
 
     def encoder_room(self, clip, tokens, fallback_peak):
diff --git a/freevideo_engine/resident_process.py b/freevideo_engine/resident_process.py
index 6003da8f..8337bdce 100644
--- a/freevideo_engine/resident_process.py
+++ b/freevideo_engine/resident_process.py
@@ -23,17 +23,17 @@
 
 ENV = 'FREEVIDEO_RESIDENT_SESSION'
 GiB = 2**30
-# A finished request keeps the worker for the next one while host memory is ample
-# (for as long as nothing needs the memory, unless KEEP_IDLE_SECONDS sets a limit), and it is given up at once when anything else
-# needs the memory: another program's GPU use grows, host memory runs low, an
+# A successful request settles individual caches inside the worker, then keeps
+# it without an idle limit unless KEEP_IDLE_SECONDS sets one. It gives up the
+# worker at once when another program's GPU use grows, host memory runs low, an
 # image or prompt model starts (those release it themselves), or the agent asks
-# for the card back. A request that is not followed by a queued part and fails,
-# or a machine without the room, releases it as before; a next part already on
-# its way is held for HOLD_SECONDS.
+# for the card back. A failed request without a queued part releases it as
+# before; a next part already on its way is held for HOLD_SECONDS.
 HOLD_SECONDS = 90.
 AGENT_HOLD_SECONDS = 30.
 KEEP_IDLE_SECONDS = None
 OTHER_GPU_GROWTH = 512 * 2**20
+PRELOAD_CONTEXT_BYTES = 1536 * 2**20  # what the preload's process adds beside its weights (context, allocator, staging; measured 1.4 GiB)
 HOST_FLOOR = 4 * GiB
 # Below this much host memory (free plus the idle worker's own), a kept worker's
 # base leaves the next part less room than a fresh one; measured on 16 GiB.
@@ -363,6 +363,7 @@ def __init__(self):
         self.identity = None
         self.active = False
         self.prewarm = None
+        self.gpu = self.baseline = self.own = None
         self.mutex = threading.RLock()
         self.generation = 0
         self.idle_since = None
@@ -384,7 +385,7 @@ def take_decisions(self):
     def finished(self, *, next_part=None, agent=False, gpu=None, warm=False):
         """Release the idle worker now, or keep it for the next request.
 
-        A successful request (warm) keeps it while host memory is ample and the card's use can be read, so
+        A successful request (warm) keeps its settled caches above HOST_FLOOR while the card's use can be read, so
         another program's growth is seen; a next part on its way keeps it for a short hold. Anything else
         gives the memory back now.
         """
@@ -398,6 +399,8 @@ def finished(self, *, next_part=None, agent=False, gpu=None, warm=False):
                 warm = False
             if not next_part or (hold is not None and hold <= 0):
                 return self.release('finished' if not next_part else 'no_hold')
+            # The worker settles individual caches against its live RAM budget.
+            # A small host alone is no reason to discard everything it kept.
             reason = pressure(gpu, None) or self._tight()
             if reason:
                 return self.release(reason)
@@ -406,21 +409,12 @@ def finished(self, *, next_part=None, agent=False, gpu=None, warm=False):
             baseline = gpu_used(gpu)
             if warm and baseline is None:
                 return self.release('gpu_unreadable')
-            threading.Thread(target=self._watch, args=(self.generation, self.idle_since, hold, gpu, baseline),
+            self.gpu, self.baseline = gpu, baseline
+            self.own = gpu_own(gpu) if warm else None
+            threading.Thread(target=self._watch, args=(self.generation, self.idle_since, hold, gpu),
                              name='freevideo-resident-hold', daemon=True).start()
             return self.record(action='kept', reason=next_part, seconds=hold, pid=self.process.pid)
 
-    def release_idle(self, reason='requested'):
-        """Give the card and host memory back on a caller's request; never touches a running request."""
-        with self.mutex:
-            if self.active:
-                return dict(released=False, reason='busy')
-            if self.process is None or self.process.poll() is not None:
-                return dict(released=False, reason='nothing_held')
-            row = self.release(reason)
-        return dict(released=True, reason=reason, **{k: v for k, v in (row or {}).items()
-                                                      if k in ('idle_seconds', 'worker_rss_bytes')})
-
     def _tight(self):
         from .system import windows
         if windows():
@@ -436,11 +430,79 @@ def _tight(self):
             pass
         return None
 
-    def _watch(self, generation, since, hold, gpu, baseline):
+    def preloading(self):
+        return self.prewarm is not None and self.prewarm.thread is not None and self.prewarm.thread.is_alive()
+
+    def warm_encoder(self, output, python, environment, *, busy=None, notify=None):
+        """Place the encoder on the card while the kept worker idles, under the worker's own lease.
+
+        Only the models the worker can hold beside what it keeps are placed (it
+        sizes that from live free memory), and any new request, queued prompt or
+        memory pressure stops it first.
+        """
+        with self.mutex:
+            if (self.process is None or self.process.poll() is not None or self.active or
+                    environment.get('FREEVIDEO_ENCODER_PREWARM', 'auto').lower() in ('0', 'off', 'false')):
+                return False
+            self.stop_prewarm()
+            if self.process is None:
+                return False
+            from .idle_encoder import IdleEncoder
+            self.prewarm = IdleEncoder()
+            return self.prewarm.schedule(self.endpoint, output, python, environment, busy=busy, notify=notify)
+
+    def release_idle(self, reason='requested'):
+        """Give the card and host memory back on a caller's request; never touches a running request."""
+        with self.mutex:
+            if self.active:
+                return dict(released=False, reason='busy')
+            if self.process is None or self.process.poll() is not None:
+                return dict(released=False, reason='nothing_held')
+            row = self.release(reason)
+        return dict(released=True, reason=reason, **{k: v for k, v in (row or {}).items()
+                                                      if k in ('idle_seconds', 'worker_rss_bytes')})
+
+    def placed_bytes(self):
+        """Card bytes the last idle preload accounts for: its reported growth, else its budget, plus the context."""
+        try:
+            report = json.loads(Path(self.prewarm.report).read_text(encoding='utf-8'))
+            before, after = report.get('gpu_bytes_before'), report.get('gpu_bytes_after')
+            if type(before) is int and type(after) is int:
+                known = max(0, after - before)
+            else:
+                known = max(0, int(report.get('gpu_budget_bytes') or 0))
+        except (AttributeError, OSError, TypeError, ValueError):
+            known = 0
+        return known + PRELOAD_CONTEXT_BYTES
+
+    def _watch(self, generation, since, hold, gpu):
+        placing = False
         while True:
             time.sleep(.5)
-            reason = ('idle' if hold is not None and time.monotonic() - since >= hold
-                      else pressure(gpu, baseline))
+            # Our own idle preload grows the card; only another program's growth counts,
+            # so the reference is read again once placement has ended.
+            unreadable = False
+            if self.preloading():
+                placing = True
+            elif placing:
+                placing = False
+                used = gpu_used(gpu)
+                own = gpu_own(gpu)
+                if used is None:
+                    unreadable = True  # Same as at the start of the hold: never keep without a reading.
+                elif own is not None and self.own is not None and self.baseline is not None:
+                    # Only the part this tree added moves the reference; another program's share stays visible.
+                    self.baseline += max(0, own - self.own)
+                    self.own = own
+                elif self.baseline is not None:
+                    # No per-process bytes (WDDM): only what the preload reports it placed, plus its
+                    # context, moves the reference; the rest of the growth is another program's.
+                    self.baseline += min(max(0, used - self.baseline), self.placed_bytes())
+                else:
+                    self.baseline = used
+            reason = ('gpu_unreadable' if unreadable else
+                      'idle' if hold is not None and time.monotonic() - since >= hold
+                      else pressure(gpu, None if placing else self.baseline))
             with self.mutex:
                 if generation != self.generation or self.active or self.process is None:
                     return
@@ -564,6 +626,23 @@ def gpu_used(gpu):
         return None
 
 
+def gpu_own(gpu):
+    """Card bytes held by this process tree (the kept worker and its leases), or None if not listed."""
+    try:
+        from .background_room import Gpu, family
+        device = Gpu(gpu)
+        try:
+            rows = device.processes()
+        finally:
+            device.close()
+        if rows is None or any(size is None for size in rows.values()):
+            return None
+        pids = family(os.getpid())
+        return sum(size for pid, size in rows.items() if pid in pids)
+    except (OSError, RuntimeError, AttributeError, ValueError, ImportError):
+        return None
+
+
 def pressure(gpu, baseline):
     """Why a kept idle worker should go now, or None."""
     try:
diff --git a/freevideo_engine/resident_worker.py b/freevideo_engine/resident_worker.py
index 14192674..bd1b01a0 100644
--- a/freevideo_engine/resident_worker.py
+++ b/freevideo_engine/resident_worker.py
@@ -68,6 +68,25 @@ def run(message):
                         from .worker import generate
                         bank.configure(request['gpu_budget_bytes'], request.get('ram_budget_bytes'))
                         generate(request, resident=bank)
+                        # Only a fully returned video is a success, including its
+                        # final cleanup. Keep habit data inside this worker.
+                        try:
+                            bank.succeeded(request)
+                        except Exception:
+                            pass  # Optional observations cannot fail a finished video.
+                        try:
+                            bank.settle(float(request['engine_options'].get('pin_host_gb') or 0)*1e9)
+                        except Exception:
+                            pass  # Best-effort trimming cannot fail a finished video.
+                        if request.get('metrics'):
+                            try:
+                                from .monitoring import save
+                                path = Path(request['metrics'])
+                                metrics = json.loads(path.read_text(encoding='utf-8'))
+                                metrics.update(resident_models=bank.snapshot(), resident_admission=list(bank.decisions))
+                                save(path, metrics)
+                            except (OSError, ValueError, TypeError):
+                                pass
                     else:
                         from .encode_worker import encode
                         bank.configure(request['gpu_budget_gb']*1e9, request.get('ram_budget_bytes'))
@@ -79,7 +98,7 @@ def run(message):
                 if diagnostics is not None:
                     diagnostics.finish(False)
                 traceback.print_exc()
-                from .adaptive import classify_failure, commit_exhausted, cuda_runtime_error
+                from .adaptive import classify_failure, commit_exhausted, commit_limited_allocation, cuda_runtime_error
                 failure = classify_failure(error)
                 if request.get('metrics'):
                     from .monitoring import save
@@ -119,6 +138,12 @@ def run(message):
                 if commit_exhausted(error):
                     log.flush()
                     os._exit(75)
+                # The same holds when the driver refuses a device allocation
+                # for lack of commit while the card still has free memory: the
+                # process keeps its charge and its mappings are unreliable.
+                if commit_limited_allocation(error, failure):
+                    log.flush()
+                    os._exit(75)
                 failed_error = error
 
             # ModelBank cannot own tensors held by an exception's traceback.
diff --git a/freevideo_engine/runtime_libraries.py b/freevideo_engine/runtime_libraries.py
index 2fc9e804..bb3d43a6 100644
--- a/freevideo_engine/runtime_libraries.py
+++ b/freevideo_engine/runtime_libraries.py
@@ -167,6 +167,10 @@ def diagnose(site=None):
         result.update(found)
     except Exception as error:
         result['check_error'] = '%s: %s' % (type(error).__name__, error)
+    try:
+        result['missing_files'] = {name: files for name, files in list(package_files(site).items())[:6]}
+    except Exception as error:
+        result['missing_files_error'] = '%s: %s' % (type(error).__name__, error)
     try:
         for path in loaded():
             inside = _canonical(os.path.dirname(path)) in directories
@@ -187,6 +191,13 @@ def summary(result, repair=REINSTALL):
     if changed:
         lines.append("Some of FreeVideo's PyTorch library files do not match the installed version (%s). %s"
                      % (', '.join(changed[:4]) + (', …' if len(changed) > 4 else ''), repair))
+    missing = result.get('missing_files')
+    if isinstance(missing, dict) and missing:
+        names = ['%s (%s)' % (name, ', '.join(str(file).rsplit('/', 1)[-1] for file in files[:2]))
+                 for name, files in list(missing.items())[:4] if isinstance(files, list) and files]
+        if names:
+            lines.append('Files of these Python packages are missing or have changed: %s%s. %s'
+                         % (', '.join(names), ', …' if len(missing) > 4 else '', repair))
     outside = sorted({str(Path(p).parent) for p in result.get('outside', [])})
     if outside:
         lines.append('cuDNN was also loaded from outside FreeVideo: %s.' % '; '.join(outside[:3]))
@@ -363,6 +374,83 @@ def incomplete(python):
         return []
 
 
+def package_files(site, limit=3):
+    """Files RECORD lists inside site-packages that are missing or have another size.
+
+    Returns {distribution: [paths relative to site-packages, at most `limit`]}. Only
+    files below site-packages count: console scripts and data files sit beside it, and
+    their size depends on where the environment was created. A numpy extension that
+    cannot load is usually one of these files gone (quarantined, cleaned up, cut short),
+    which an ordinary install does not notice. No hashing, so it is cheap to run.
+    Unreadable records and checks that cannot run find nothing.
+    """
+    found = {}
+    try:
+        site = Path(site)
+        # Paths stay as RECORD spells them: a package folder that is a junction
+        # to another drive is still checked.
+        base = os.path.normcase(os.path.normpath(str(site))) + os.sep
+        for distribution in importlib.metadata.distributions(path=[str(io_path(site))]):
+            try:
+                rows = list(csv.reader(io.StringIO(distribution.read_text('RECORD') or '')))
+                name = distribution.metadata['Name']
+            except (OSError, csv.Error, UnicodeDecodeError, KeyError):
+                continue
+            if not name:
+                continue
+            for row in rows:
+                if len(row) < 3 or not row[2].isdigit():
+                    continue
+                path = os.path.normpath(os.path.join(str(site), row[0]))
+                if not os.path.normcase(path).startswith(base):
+                    continue
+                try:
+                    size = os.path.getsize(io_path(path))
+                except (FileNotFoundError, NotADirectoryError):
+                    size = None
+                except OSError:
+                    continue
+                if size != int(row[2]):
+                    files = found.setdefault(name, [])
+                    if len(files) >= limit:
+                        break
+                    files.append(row[0].replace('\\', '/'))
+    except Exception:
+        return {}
+    return found
+
+
+def gpu_family(name):
+    """The packages installed from the GPU wheels: PyTorch, its CUDA libraries and its compiler."""
+    name = normalize_name(name)
+    return name in ('torch', 'torchvision', 'torchaudio', 'triton', 'triton-windows') or name.startswith('nvidia-')
+
+
+def pinned_names(constraints):
+    """Normalized names a constraints file pins, or None when it cannot be read."""
+    rows = _requirement_rows(constraints)
+    if rows is None:
+        return None
+    return {normalize_name(match[0]) for match in (re.match(r'[A-Za-z0-9][A-Za-z0-9._-]*', row) for row in rows) if match}
+
+
+def missing_packages(python, constraints):
+    """Pinned distributions of this environment whose RECORD files are missing or cut short.
+
+    Setup and repair reinstall these: uv treats a package at the pinned version
+    as installed whatever its files look like. Packages the constraints do not
+    pin (editable installs, kernels built on this machine) are left alone.
+    """
+    try:
+        site = site_for(python)
+        pinned = pinned_names(constraints)
+        if site is None or pinned is None:
+            return []
+        return sorted(name for name in package_files(site, limit=1) if normalize_name(name) in pinned)
+    except Exception:
+        return []
+
+
 def installed_bytes(python, name):
     """Installed size of one distribution in this Python's environment, from its RECORD."""
     try:
diff --git a/freevideo_engine/usage_habit.py b/freevideo_engine/usage_habit.py
new file mode 100644
index 00000000..b7f1154b
--- /dev/null
+++ b/freevideo_engine/usage_habit.py
@@ -0,0 +1,116 @@
+"""Recent usage habit kept across restarts: operation types and counts only.
+
+Nothing derived from a prompt, image or seed is stored. A transition between
+two successive successful videos is one of three kinds of change, and the file
+keeps the last three transitions (the four-request window) plus running counts.
+A file that is missing, unreadable or malformed counts as no habit at all.
+
+The same file keeps the last OPS_STORED operations ('image' or 'video') under "ops", which the image
+worker reads to decide how long to stay on the card. Reading and saving the habit leave them as they are.
+"""
+import json
+import os
+from pathlib import Path
+
+WINDOW = 4
+KINDS = ('images', 'prompt', 'rerun')
+OPS = ('image', 'video')
+OPS_STORED = 8
+
+
+def path(root=None):
+    override = os.environ.get('FREEVIDEO_USAGE_HABIT')
+    if override:
+        return Path(override)
+    if root is None:
+        from .system import install_root
+        root = install_root()
+    return Path(root) / '.freevideo' / 'usage-habit.json'
+
+
+def transition(previous, current):
+    """Kinds of change between two requests' digests (never stored)."""
+    kinds = []
+    if previous['images'] != current['images']:
+        kinds.append('images')
+    if previous['prompt'] != current['prompt']:
+        kinds.append('prompt')
+    elif not kinds:
+        kinds.append('rerun')
+    return kinds
+
+
+def transitions(history):
+    recent = history[-WINDOW:]
+    return [transition(a, b) for a, b in zip(recent, recent[1:])]
+
+
+def dominant(recent):
+    """The kind that two thirds of the recent transitions share, or None.
+
+    Image changes win ties with prompt changes, then same-prompt reruns.
+    """
+    if not recent:
+        return None
+    for kind in KINDS:
+        if sum(kind in kinds for kinds in recent) * 3 >= len(recent) * 2:
+            return kind
+    return None
+
+
+def load(file=None):
+    """The stored habit, or None when there is nothing trustworthy to read."""
+    try:
+        value = json.loads(Path(file or path()).read_text(encoding='utf-8'))
+        if value['version'] != 1 or value['window'] != WINDOW:
+            return None
+        recent = value['recent']
+        counts = value['counts']
+        if (not isinstance(recent, list) or len(recent) > WINDOW - 1
+                or any(not isinstance(kinds, list) or any(kind not in KINDS for kind in kinds) for kinds in recent)
+                or not isinstance(counts, dict)
+                or any(type(counts.get(name)) is not int or counts[name] < 0 for name in (*KINDS, 'requests'))):
+            return None
+        ops = value.get('ops', [])
+        if not isinstance(ops, list) or len(ops) > OPS_STORED or any(kind not in OPS for kind in ops):
+            ops = []
+        return dict(version=1, window=WINDOW, recent=[list(kinds) for kinds in recent],
+                    counts={name: counts[name] for name in (*KINDS, 'requests')}, ops=list(ops))
+    except (OSError, ValueError, KeyError, TypeError, AttributeError):
+        return None
+
+
+def empty():
+    return dict(version=1, window=WINDOW, recent=[], counts=dict.fromkeys((*KINDS, 'requests'), 0), ops=[])
+
+
+def record(kinds, file=None):
+    """Append one transition (None for the first request after a restart) and save."""
+    from .monitoring import save
+    value = load(file) or empty()
+    value['counts']['requests'] += 1
+    if kinds is not None:
+        value['recent'] = (value['recent'] + [list(kinds)])[-(WINDOW - 1):]
+        for kind in kinds:
+            value['counts'][kind] += 1
+    save(Path(file or path()), value if value['ops'] else {name: item for name, item in value.items() if name != 'ops'})
+    return value
+
+
+def ops(file=None):
+    """The last operations ('image' or 'video'), oldest first; empty when nothing can be trusted."""
+    value = load(file)
+    return value['ops'] if value else []
+
+
+def note_op(kind, file=None):
+    """Remember one operation next to the habit. Only the kind is stored, nothing from the request."""
+    if kind not in OPS:
+        raise ValueError(kind)
+    from .monitoring import save
+    value = load(file) or empty()
+    value['ops'] = (value['ops'] + [kind])[-OPS_STORED:]
+    try:
+        save(Path(file or path()), value)
+    except OSError:
+        pass  # A habit that cannot be saved is only forgotten at restart.
diff --git a/freevideo_engine/worker.py b/freevideo_engine/worker.py
index 5a365d58..ea7fa7b4 100644
--- a/freevideo_engine/worker.py
+++ b/freevideo_engine/worker.py
@@ -763,12 +763,6 @@ def keep_audio_lead_in(value):
             resident.device_free, resident.device_reserve = None, 0
         if live_budget is not None:
             live_budget.close()
-        if resident is not None and metrics['success'] and engine is not None:
-            try:
-                resident.settle(float(request['engine_options'].get('pin_host_gb') or 0)*1e9)
-            except Exception:
-                # Trimming is best effort; it must not turn a finished request into a failure.
-                pass
         if resident is not None:
             metrics['resident_admission'] = list(resident.decisions)
         from .adaptive import classify_failure
diff --git a/freevideo_source_bootstrap.py b/freevideo_source_bootstrap.py
index 8aca0723..b05644f3 100644
--- a/freevideo_source_bootstrap.py
+++ b/freevideo_source_bootstrap.py
@@ -101,6 +101,10 @@ def root():
     'web/assets/tour/c8-s1.webp',
     'web/assets/tour/c9-s1.mp4',
     'web/assets/tour/c9-s1.webp',
+    'web/assets/tour/c10-person.webp',
+    'web/assets/tour/c10-s1.mp4',
+    'web/assets/tour/c10-s1.webp',
+    'web/assets/tour/c10-source.webp',
     'web/assets/tour/fox-sheet.webp',
 ))
 
diff --git a/web/assets/tour/c10-person.webp b/web/assets/tour/c10-person.webp
new file mode 100644
index 00000000..14567608
Binary files /dev/null and b/web/assets/tour/c10-person.webp differ
diff --git a/web/assets/tour/c10-s1.mp4 b/web/assets/tour/c10-s1.mp4
new file mode 100644
index 00000000..35d08cda
Binary files /dev/null and b/web/assets/tour/c10-s1.mp4 differ
diff --git a/web/assets/tour/c10-s1.webp b/web/assets/tour/c10-s1.webp
new file mode 100644
index 00000000..4d9c4db1
Binary files /dev/null and b/web/assets/tour/c10-s1.webp differ
diff --git a/web/assets/tour/c10-source.webp b/web/assets/tour/c10-source.webp
new file mode 100644
index 00000000..87c19255
Binary files /dev/null and b/web/assets/tour/c10-source.webp differ
diff --git a/web/cast_generate.css b/web/cast_generate.css
index bfd453b7..828c679a 100644
--- a/web/cast_generate.css
+++ b/web/cast_generate.css
@@ -2,6 +2,11 @@
 .fv-cg-body{display:grid;grid-template-columns:minmax(0,2fr) minmax(0,3fr);gap:24px;align-items:start}
 .fv-cg [hidden]{display:none!important}
 .fv-cg-side{display:flex;flex-direction:column;gap:12px;min-width:0;align-self:stretch}
+/* The LoRA list opens in the column, under the description, so it never leaves the dialog; its controls take the panel's look. */
+.fv-cg-side .fv-ie-lora summary{display:inline-grid;box-sizing:border-box;height:32px;padding:0 9px;border:1px solid var(--fv-border);border-radius:var(--fv-r-sm);white-space:nowrap;font-variant-numeric:tabular-nums}
+.fv-cg-side .fv-ie-lora-trigger::placeholder{color:var(--fv-disabled)}
+.fv-cg-side .fv-ie-lora .fv-ie-lora-panel{position:static;width:auto;max-height:none;margin-top:8px;padding:0;border:0;background:none;box-shadow:none}
+.fv-cg-side .fv-ie-lora select,.fv-cg-side .fv-ie-lora input:not([type=checkbox]),.fv-cg-side .fv-ie-lora button{box-sizing:border-box;min-height:30px;padding:3px 8px;border:1px solid var(--fv-border,#ffffff1f);border-radius:6px;background:var(--fv-raised,#202a38);color:inherit;font:inherit}
 .fv-cg-foot{display:grid;grid-template-columns:auto minmax(0,1fr) auto;align-items:center;gap:8px}
 .fv-cg-text{flex:1;box-sizing:border-box;width:100%;min-height:160px;resize:none;padding:12px;border:1px solid var(--fv-border,#ffffff15);border-radius:8px;background:var(--fv-card,#202a38);color:inherit;font:inherit;line-height:1.6;outline:none;transition:border-color .2s}
 .fv-cg-text:focus{border-color:color-mix(in srgb,var(--fv-accent,#6db8fa) 60%,transparent)}
diff --git a/web/cast_generate.js b/web/cast_generate.js
index cc2543f4..99347dc6 100644
--- a/web/cast_generate.js
+++ b/web/cast_generate.js
@@ -4,6 +4,7 @@
 // and stay in the input folder; the dialog puts them on the card.
 import {createImageGenClient, imageGenProgress, imagePackUnavailableMessage, renderImageFailure,
     IMAGE_GEN_POLL_MS, IMAGE_GEN_INFO_POLL_MS} from './image_generation.js';
+import {createImageLoras} from './image_lora.js';
 
 export {imageGenerationSupported as castGenerateSupported} from './image_generation.js';
 
@@ -22,7 +23,7 @@ const randomSeed = () => crypto.getRandomValues(new Uint32Array(1))[0] >>> 1;
 export function createCastGenerator({client, onChange = () => {}, schedule = setTimeout, unschedule = clearTimeout}) {
     const fresh = () => CAST_VIEWS.map(view => ({view, file: '', fraction: 0}));
     let state = {pictures: fresh(), running: false, current: -1, error: '', report_id: null, waiting: '', description: ''};
-    let epoch = 0, timer = null, job = null, plan = null, seed = 0, disposed = false;
+    let epoch = 0, timer = null, job = null, plan = null, seed = 0, disposed = false, loras;
     const emit = patch => { state = {...state, ...patch}; if (!disposed) onChange(state); };
     const setPicture = (index, patch) => emit({pictures: state.pictures.map((row, i) => i === index ? {...row, ...patch} : row)});
     const cancelJob = id => Promise.resolve().then(() => client.cancel(id)).catch(() => {});
@@ -31,7 +32,8 @@ export function createCastGenerator({client, onChange = () => {}, schedule = set
     async function runOne(index, alive) {
         const view = plan[index], front = state.pictures[0].file;
         const row = await client.generate({prompt: view.prompt, width: view.width, height: view.height, seed,
-            images: view.from_front ? [front] : [], ...(view.view === 'back' ? {cast_view: 'back'} : {})});
+            images: view.from_front ? [front] : [], ...(view.view === 'back' ? {cast_view: 'back'} : {}),
+            ...(loras === undefined ? {} : {loras})});
         if (!row.job) throw Object.assign(new Error(row.error || 'generation_failed'), {code: row.error, report_id: row.report_id, memory: row.memory});
         if (!alive()) { await cancelJob(row.job); return null; }
         job = row.job;
@@ -54,14 +56,15 @@ export function createCastGenerator({client, onChange = () => {}, schedule = set
         }
     }
 
-    async function start(description) {
+    async function start(description, {loras: selectedLoras} = {}) {
         description = String(description || '').trim();
         if (state.running || disposed || !description) return;
         // After a failure the same description goes on from the picture that failed.
-        const resume = !!state.error && description === state.description && plan && !done();
+        const resume = !!state.error && description === state.description && plan && !done()
+            && JSON.stringify(selectedLoras) === JSON.stringify(loras);
         const current = ++epoch;
         const alive = () => current === epoch && !disposed;
-        if (!resume) { seed = randomSeed(); plan = null; }
+        if (!resume) { seed = randomSeed(); plan = null; loras = structuredClone(selectedLoras); }
         emit({running: true, error: '', report_id: null, waiting: '', description,
             ...(resume ? {} : {pictures: fresh(), current: 0})});
         try {
@@ -105,7 +108,7 @@ export function createCastGenerator({client, onChange = () => {}, schedule = set
         dispose() { disposed = true; return stop(); }};
 }
 
-export function mountCastGenerate({container, api, text, t = text, viewURL, onState = () => {}, fetchApi, client}) {
+export function mountCastGenerate({container, api, text, t = text, viewURL, onState = () => {}, fetchApi, client, loras}) {
     const tr = t || ((en) => en);
     const fetcher = fetchApi || (api ? api.fetchApi.bind(api) : globalThis.fetch.bind(globalThis));
     client = client || createImageGenClient(fetcher);
@@ -129,8 +132,11 @@ export function mountCastGenerate({container, api, text, t = text, viewURL, onSt
     field.maxLength = 1000;
     const note = el('div', 'fv-cg-note'); note.setAttribute('role', 'status');
     const action = el('button', 'fv-cg-action'); action.type = 'button';
+    // A caller that passes a fixed list owns the choice; otherwise the panel offers the same picker as image generation.
+    const picker = loras === undefined ? createImageLoras({t: tr, fetchApi: fetcher, getPrompt: () => field.value,
+        setPrompt: text => { field.value = text; }, onChange: () => render()}) : null;
     foot.append(note, action);
-    side.append(field, foot);
+    side.append(field, ...(picker ? [picker.element] : []), foot);
     const grid = el('div', 'fv-cg-views');
     const cells = CAST_VIEWS.map(view => {
         const cell = el('div', 'fv-cg-view');
@@ -157,6 +163,7 @@ export function mountCastGenerate({container, api, text, t = text, viewURL, onSt
         const isMissing = missing();
         root.dataset.state = isMissing ? 'missing' : running ? 'running' : generator.done() ? 'done' : state.error ? 'failed' : 'idle';
         field.readOnly = running;
+        if (picker) picker.element.inert = running;
         action.hidden = !!isMissing;
         // One accent button at a time: Generate until the five pictures exist, then Regenerate is neutral.
         const finishedAll = generator.done() && !running;
@@ -212,7 +219,8 @@ export function mountCastGenerate({container, api, text, t = text, viewURL, onSt
             return;
         }
         attention = null;
-        void generator.start(field.value);
+        const chosen = picker ? picker.value() : loras;
+        void generator.start(field.value, picker && !chosen.length ? {} : {loras: chosen});
     }
     action.onclick = () => generator.state().running ? void generator.stop() : begin();
     field.addEventListener('input', () => {
@@ -230,7 +238,7 @@ export function mountCastGenerate({container, api, text, t = text, viewURL, onSt
         getPictures: () => generator.done() && !generator.state().running ? generator.state().pictures.map(row => ({file: row.file, label: row.view})) : [],
         isBusy: () => generator.state().running,
         cancel: () => generator.stop(),
-        destroy() { disposed = true; clearTimeout(infoTimer); void generator.dispose(); root.remove(); },
+        destroy() { disposed = true; picker?.dispose(); clearTimeout(infoTimer); void generator.dispose(); root.remove(); },
         element: root,
     };
 }
diff --git a/web/feature_tour_content.js b/web/feature_tour_content.js
index 4fb21d46..25025416 100644
--- a/web/feature_tour_content.js
+++ b/web/feature_tour_content.js
@@ -16,6 +16,8 @@ export const FEATURE_TOUR = [
     card('reference-image', copy('Reference image', '参考图'), copy('The same character stays consistent across scenes.', '换了场景，角色仍是同一个。'),
         [{src: 'fox-sheet.webp'}, {text: copy('A fox walks through a forest.', '狐狸走过森林。')}, {text: copy('A fox walks through snow.', '狐狸走过雪地。')}], clip('c5-forest-s1'),
         {assets: [{file: 'fox-sheet.webp', role: 'reference'}], prompt: ['The fox in {image1} walks through a sunlit forest.', '{image1} 中的狐狸走过洒满阳光的森林。'], settings: {}}, {sequence: [clip('c5-forest-s1'), clip('c5-snow-s1')]}),
+    card('replace-person', copy('Replace person', '替换人物'), copy('Put the person from an image into a video.', '把视频里的人换成图中的人物。'),
+        [{src: 'c10-source.webp'}, {src: 'c10-person.webp'}], clip('c10-s1'), {status: 'pending-assets'}, {trigger: 'media', added: '0.4.28'}),
     card('extend', copy('Extend', '接着生成'), copy('Continue from where the last clip ends.', '从上一段的结尾接着生成下一段。'),
         [{src: 'c6-s1.webp'}], clip('c6-s1'), {status: 'pending-assets', view: 'editor'}, {trigger: 'editor', previewRange: [2, 8]}),
     card('two-pass', copy('Two-pass', '二次采样'), copy('The first pass sets the composition; the second adds detail.', '第一遍定构图，第二遍补细节。'),
diff --git a/web/health.js b/web/health.js
index 0e2347db..3b1a292b 100644
--- a/web/health.js
+++ b/web/health.js
@@ -15,7 +15,7 @@ const launchVisit = query.get('freevideo') === 'launch';
 // Every script this folder ships; a test keeps the list in step with web/*.js.
 export const MODULES = ['audio_original.js', 'chat_data.js', 'phone_model.js', 'phone.js', 'lan_access.js', 'qrcodegen.js', 'agent_activity.js', 'agent_confirm_settings.js', 'agent_connect.js', 'agent_session.js', 'autopilot.js', 'autopilot_chat.js', 'chat_attachments.js', 'autopilot_conversations.js', 'browser_extension.js', 'caret_rect.js', 'chain_support.js', 'chain_rules.js', 'capability_view.js', 'cast_generate.js', 'cast_mention.js', 'cast_overlay.js', 'cast_source.js', 'compatibility.js', 'entry_showcase.js', 'error_panel.js', 'first_run_guide.js', 'fonts.js', 'freevideo.js', 'generation_progress.js',
     'editor.js', 'editor_actions.js', 'editor_lifecycle.js', 'editor_adjust.js', 'editor_adjust_picture.js', 'editor_picture_fields.js', 'editor_picture_session.js', 'editor_preview_picture.js', 'editor_entry.js', 'editor_expand.js', 'editor_export_dialog.js', 'editor_first_frame.js', 'editor_media_references.js', 'editor_references.js', 'editor_genbar.js', 'editor_geometry.js', 'editor_library.js', 'editor_menu.js', 'editor_model.js', 'editor_frame.js', 'editor_player.js', 'editor_room.js', 'editor_select.js', 'editor_state.js', 'editor_store.js', 'editor_tracks.js', 'editor_generation.js', 'editor_transport.js', 'editor_transport_model.js', 'editor_audio.js', 'editor_audio_model.js', 'editor_audio_notes.js', 'editor_audio_replace.js', 'editor_audio_rules.js', 'editor_speed.js', 'editor_speed_model.js', 'editor_failure.js', 'editor_header_fit.js',
-    'editor_loader.js', 'editor_text_coverage.js', 'editor_text.js', 'editor_toolbar_more.js', 'editor_text_model.js', 'editor_text_layout.js', 'editor_notes.js', 'editor_icons.js', 'dialog_keys.js', 'done_notify.js', 'failure_issue.js', 'feedback.js', 'frame_sizes.js', 'health.js', 'image_commands.js', 'image_editor.js', 'image_editor_engine.js', 'image_generation.js', 'launch_visit.js', 'launcher.js', 'length_plan.js', 'lora_triggers.js', 'lora_check.js', 'library.js', 'motion.js', 'native_errors.js', 'translate_guard.js', 'output_download.js', 'port_notice.js', 'preview_scene.js',
+    'editor_loader.js', 'editor_text_coverage.js', 'editor_text.js', 'editor_toolbar_more.js', 'editor_text_model.js', 'editor_text_layout.js', 'editor_notes.js', 'editor_icons.js', 'dialog_keys.js', 'done_notify.js', 'failure_issue.js', 'feedback.js', 'frame_sizes.js', 'health.js', 'image_commands.js', 'image_editor.js', 'image_editor_engine.js', 'image_generation.js', 'image_lora.js', 'launch_visit.js', 'launcher.js', 'length_plan.js', 'lora_triggers.js', 'lora_check.js', 'library.js', 'motion.js', 'native_errors.js', 'translate_guard.js', 'output_download.js', 'port_notice.js', 'preview_scene.js',
     'progress_connection.js', 'prompt_draft.js', 'prompt_enhance.js', 'prompt_references.js', 'prompt_sections.js', 'prompt_terms.js', 'request_assets.js', 'repair_hint.js', 'report_issue.js', 'issue_button.js', 'studio_recent.js',
     'result_actions.js', 'result_reuse.js', 'sampling_effort.js', 'setup.js', 'share.js', 'share_layout.js', 'studio.js', 'studio_chain.js', 'studio_extend.js', 'studio_length.js', 'first_frame_row.js', 'replace_person.js', 'replace_state.js', 'optional_inputs.js', 'studio_queue.js', 'studio_size.js',
     'toolbar_position.js', 'updates.js', 'view_navigation.js', 'wordmark.js', 'workspace_restore.js', 'cast_library.js', 'media_upload.js', 'whitebox_encode.js', 'whitebox_mp4.js', 'whitebox_recording.js', 'whitebox_render.js', 'whitebox_scenes.js', 'whitebox_worker.js',
diff --git a/web/image_editor.css b/web/image_editor.css
index ecb2b5bf..5207ca64 100644
--- a/web/image_editor.css
+++ b/web/image_editor.css
@@ -245,6 +245,25 @@ button.fv-ie-size:hover{background:var(--fv-hover);color:var(--fv-ink)}
 .fv-ie-menu summary{display:grid;place-items:center;border-radius:var(--fv-r-sm);cursor:pointer;list-style:none}
 .fv-ie-menu summary::-webkit-details-marker{display:none}
 .fv-ie-menu-list{position:absolute;right:0;bottom:calc(100% + 6px);z-index:3;width:236px;max-height:300px;overflow:auto;padding:6px;border:1px solid var(--fv-border);border-radius:var(--fv-r-md);background:var(--fv-card);box-shadow:0 14px 40px rgba(0,0,0,.5);display:flex;flex-direction:column;gap:2px}
+.fv-ie-lora summary{width:auto;padding:0 9px;border:1px solid var(--fv-border);border-radius:var(--fv-r-sm);height:32px;white-space:nowrap;font-variant-numeric:tabular-nums}
+.fv-ie-lora[data-active=true] summary{color:var(--fv-ink);border-color:var(--fv-accent-dim)}
+.fv-ie-lora .fv-ie-lora-panel{top:calc(100% + 6px);right:0;bottom:auto;width:min(340px,calc(100vw - 24px));max-height:360px;gap:8px;padding:10px}
+@media (max-width:700px){.fv-ie-lora .fv-ie-lora-panel{right:auto;left:0}}
+.fv-ie-lora-list{display:flex;flex-direction:column;gap:8px}
+.fv-ie-lora-row{display:grid;grid-template-columns:18px minmax(0,1fr) 64px 28px;align-items:center;gap:6px}
+.fv-ie-lora-row select{min-width:0;width:100%;padding-right:22px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.fv-ie-lora-row input[type=number]{width:64px;padding:4px 6px;font-variant-numeric:tabular-nums}
+.fv-ie-lora-row input[type=checkbox]{accent-color:var(--fv-accent);width:15px;height:15px}
+.fv-ie-lora-row button{height:28px;padding:0 6px}
+.fv-ie-lora-row[data-enabled=false]>:not(input[type=checkbox]){opacity:.45}
+.fv-ie-lora-row .fv-ie-lora-trigger{grid-column:2/span 3;min-width:0;height:28px;box-sizing:border-box;padding:4px 8px;border:1px solid var(--fv-border);border-radius:var(--fv-r-sm);background:var(--fv-card);color:var(--fv-ink);font:inherit;font-size:var(--fv-micro)}
+.fv-ie-lora-trigger::placeholder{color:var(--fv-disabled)}
+.fv-ie-lora-problem{grid-column:2/span 3;margin:0;font-size:var(--fv-micro);color:var(--fv-danger)}
+.fv-ie-lora-problem[hidden]{display:none}
+.fv-ie-lora-row[data-unusable=true] .fv-ie-lora-trigger{display:none}.fv-ie-lora-row[data-unusable=true] input[type=number]{visibility:hidden}.fv-ie-lora-row button{justify-self:end;width:28px}
+.fv-ie-lora-actions{display:flex;justify-content:center}
+.fv-ie-lora-note{margin:0;font-size:var(--fv-micro);color:var(--fv-muted);text-align:center}
+.fv-ie-lora-note:empty{display:none}
 .fv-ie-menu-label{padding:6px 8px 2px;color:var(--fv-muted);font-size:11px}
 .fv-ie-menu-item{display:flex;align-items:center;gap:8px;width:100%;padding:6px 8px;border:0;border-radius:6px;background:transparent;text-align:left}
 .fv-ie-menu-item:hover{background:var(--fv-hover)}
diff --git a/web/image_generation.js b/web/image_generation.js
index 74f174c3..fbb81bdb 100644
--- a/web/image_generation.js
+++ b/web/image_generation.js
@@ -5,11 +5,12 @@ import {orientMatrix, rasterKinds, scaleSelection} from './image_editor_engine.j
 import {saveReport, reportIssue, showReportTip, focusReportTip} from './report_issue.js';
 import {issueButton} from './issue_button.js';
 import {about} from './request_assets.js';
+import {createImageLoras} from './image_lora.js';
 
 export const IMAGE_GEN_PROTOCOL = Object.freeze({
     root: '/freevideo/image-gen', header: 'X-FreeVideo-Image', upload: '/upload/image',
     installed: 'ready', phase: 'phase', bytes: 'bytes', downloading: 'downloading', checking: 'checking',
-    token: 'token', job: 'job', prompt: 'prompt', width: 'width', height: 'height', images: 'images', mask: 'mask', seed: 'seed',
+    token: 'token', job: 'job', prompt: 'prompt', width: 'width', height: 'height', images: 'images', mask: 'mask', seed: 'seed', loras: 'loras',
     stage: 'stage', fraction: 'fraction', file: 'file', error: 'error', name: 'name', subfolder: 'subfolder',
     states: {queued: 'waiting', starting: 'queued', progress: 'running', complete: 'done', failed: 'error',
         cancelling: 'running', cancelled: 'error'},
@@ -80,10 +81,10 @@ export function createImageGenClient(fetchImpl = globalThis.fetch.bind(globalThi
     }
     return {
         info,
-        async generate({prompt, width, height, images = [], mask, seed, cast_view}) {
+        async generate({prompt, width, height, images = [], mask, seed, loras, cast_view}) {
             const value = await post('generate', {[p.prompt]: prompt, [p.width]: width, [p.height]: height,
                 [p.images]: images, ...(mask ? {[p.mask]: mask} : {}), ...(seed === undefined ? {} : {[p.seed]: seed}),
-                ...(cast_view === 'back' ? {cast_view} : {})});
+                ...(loras === undefined ? {} : {[p.loras]: loras}), ...(cast_view === 'back' ? {cast_view} : {})});
             return {...value, job: value[p.job]};
         },
         async status(job) {
@@ -156,6 +157,12 @@ export function packView(info, t) {
 }
 export function imageGenError(code, info, t) {
     const reasons = {
+        invalid_loras: ['The LoRA selection is invalid. Choose the files again.', 'LoRA 设置有误，请重新选择文件。'],
+        invalid_lora_strength: ['LoRA strength must be between -2 and 2.', 'LoRA 强度必须在 -2 到 2 之间。'],
+        duplicate_lora: ['Select each LoRA only once. Remove the duplicate.', '同一个 LoRA 只能选择一次，请移除重复项。'],
+        too_many_loras: ['Use at most eight LoRAs per picture.', '一张图片最多使用 8 个 LoRA，请减少数量。'],
+        lora_not_found: ['The LoRA file could not be found. Choose it again.', 'LoRA 文件找不到，请重新选择。'],
+        invalid_lora_model: ['This LoRA cannot be used for pictures. Choose another one.', '此 LoRA 不适用于图片生成，请换一个。'],
         no_memory: ['Not enough GPU memory, even at the smallest size. Close other apps using the GPU, then try again.',
             '显存不够，最小尺寸也没能生成。关掉其他占用显卡的程序后再试。'],
         busy_video: ['A video is generating. Try again when it finishes.', '视频正在生成，完成后再试。'],
@@ -476,7 +483,9 @@ export function createImageGenSession({client, prepare, apply, onChange = () =>
             if (!alive()) return;
             if (state.stage === 'upload') emit({stage: ''});
             const withImages = images.length > 0;
-            const row = await client.generate({prompt: request.prompt, width: context.width, height: context.height, images, mask});
+            const row = await client.generate({prompt: request.prompt, width: context.width, height: context.height, images, mask,
+                ...(request.seed === undefined ? {} : {seed: request.seed}),
+                ...(request.loras === undefined ? {} : {loras: request.loras})});
             const {job} = row;
             if (!job) throw failure(row.error, row);
             if (row.fitted) emit({fitted: row.fitted});
@@ -548,7 +557,7 @@ export function imageJobStore(key, storage = globalThis.sessionStorage, now = Da
 
 export function createImageGeneration({client, renderer, t, getDoc, getSelection, commit, el, button, segmented,
     library = [], resolve = file => file, icon = () => '', onRatioChange, onPlace, productVersion, fetchApi, resumeKey = null}) {
-    let mode = 'edit', prompt = '', mounted = null, references = [], tier = 'standard', past = [];
+    let mode = 'edit', prompt = '', mounted = null, references = [], loras = null, tier = 'standard', past = [];
     let promptRequired = false, attentionTimer = null;
     const copies = new Map();
     // A reference is sent no larger than the largest picture the service makes.
@@ -600,7 +609,8 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
             }
             return;
         }
-        return session.start({mode, prompt, references: chosen()});
+        const chosenLoras = loras?.value() || [];
+        return session.start({mode, prompt, references: chosen(), ...(chosenLoras.length ? {loras: chosenLoras} : {})});
     }
     function drawEdited(canvas) {
         try {
@@ -779,6 +789,7 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
             b.disabled = disabled || (empty && b.dataset.value === 'edit');
             b.setAttribute('aria-pressed', String(b.dataset.value === mode));
         }); input.disabled = disabled;
+        if (loras) { if (state.running) loras.element.setAttribute('inert', ''); else loras.element.removeAttribute('inert'); }
         renderReferences(refs, disabled);
         input.placeholder = mode === 'edit' ? getSelection() ? t('What to change in the selection', '要改选区里的什么') : t('What to change', '要改什么') : t('Describe the picture', '描述画面');
         input.setAttribute('aria-label', input.placeholder);
@@ -845,7 +856,10 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
         refs.setAttribute('aria-label', t('Reference images', '参考图'));
         const box = el('div', null, 'fv-ie-gen-box');
         const input = el('textarea', null, 'fv-ie-gen-prompt'); input.rows = 1; input.maxLength = 2000; input.value = prompt;
-        box.append(input); field.append(refs, box, historyMenu());
+        box.append(input);
+        loras ||= createImageLoras({t, fetchApi, el, onChange: () => mounted && update(), getPrompt: () => prompt,
+            setPrompt: value => { prompt = value; if (mounted) { mounted.input.value = value; mounted.grow(); } }});
+        field.append(refs, box, historyMenu(), loras.element);
         const note = el('div', '', 'fv-ie-opt-note fv-ie-gen-note'); note.setAttribute('role', 'status');
         const action = button(t('Generate', '生成'), null, () => {
             const now = session.state(), pack = !now.running && now.info && !now.info.installed ? packView(now.info, t) : null;
@@ -858,10 +872,10 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
         bar.append(group);
         if (shape) bar.append(shape);
         bar.append(field, note, action, meter);
-        mounted = {bar, input, action, note, meter, fill, group, refs, ratio, size};
         const grow = () => {
             input.style.height = '0px';
             input.style.height = (prompt.trim() ? Math.min(80, Math.max(32, input.scrollHeight + 2)) : 32) + 'px';
+            box.style.height = input.style.height;
         };
         input.oninput = () => {
             prompt = input.value;
@@ -873,6 +887,7 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
                 event.preventDefault(); void start();
             }
         };
+        mounted = {bar, input, action, note, meter, fill, group, refs, ratio, size, grow};
         grow(); update();
     }
     return {mount, drawsSelection: () => mode === 'edit' && getDoc().layers.length > 0, unmount() { clearAttention(); promptRequired = false; mounted = null; },
@@ -880,5 +895,5 @@ export function createImageGeneration({client, renderer, t, getDoc, getSelection
         // The blank canvas a reloaded generation was started on, to be set again by the editor.
         resumedCanvas: resumed?.context?.mode === 'new'
             ? {width: resumed.context.doc.width, height: resumed.context.doc.height} : null,
-        dispose() { clearAttention(); return session.dispose(); }, state: session.state};
+        dispose() { clearAttention(); loras?.dispose(); return session.dispose(); }, state: session.state};
 }
diff --git a/web/image_lora.js b/web/image_lora.js
new file mode 100644
index 00000000..b036be52
--- /dev/null
+++ b/web/image_lora.js
@@ -0,0 +1,183 @@
+// A compact LoRA picker for image generation, with the video panel's rows: switch, file,
+// strength, trigger words. It keeps its own list in this browser (the image editor has no
+// graph node to hold it) and puts trigger words into the prompt the way the video panel does.
+import {reconcileLoraTriggers} from './lora_triggers.js';
+import {createVerdicts} from './lora_check.js';
+
+export const IMAGE_LORA_LIMIT = 8, IMAGE_LORA_STRENGTH = 2;
+const STORE = 'freevideo_image_loras';
+const domEl = (tag, label, cls) => { const e = document.createElement(tag); if (label != null) e.textContent = label; if (cls) e.className = cls; return e; };
+
+/** The rows worth sending: switched on, a real strength, a file that is there and fits. */
+export function activeLoras(rows, {names = null, verdicts = null} = {}) {
+    return rows.filter(row => row.enabled !== false && Number(row.strength) !== 0
+        && (!names || names.includes(row.name)) && verdicts?.peek(row.name)?.usable !== false)
+        .map(row => ({name: row.name, strength: Number(row.strength), enabled: true}));
+}
+
+export function createImageLoras({t, fetchApi, el = domEl, getPrompt = () => '', setPrompt = null, onChange = () => {}, storage = globalThis.localStorage}) {
+    let rows = [], records = [], names = null, disposed = false, skipped = 0;
+    const words = new Map(), loading = new Map(), drafts = new Map();
+    const verdicts = createVerdicts(fetchApi, {kind: 'image'});
+    try {
+        const saved = JSON.parse(storage?.getItem(STORE) || 'null');
+        if (Array.isArray(saved?.rows)) rows = saved.rows.filter(r => r && typeof r.name === 'string').slice(0, IMAGE_LORA_LIMIT)
+            .map(r => ({name: r.name, strength: Number.isFinite(Number(r.strength)) ? Number(r.strength) : 1, enabled: r.enabled !== false}));
+        // Which trigger words this panel put into a prompt is not carried over: a new description box starts unowned.
+    } catch { /* an unreadable saved list starts empty */ }
+
+    const menu = el('details', null, 'fv-ie-menu fv-ie-lora');
+    const summary = el('summary', null, 'fv-ie-lora-summary');
+    const panel = el('div', null, 'fv-ie-menu-list fv-ie-lora-panel');
+    const list = el('div', null, 'fv-ie-lora-list'), actions = el('div', null, 'fv-ie-lora-actions'), note = el('p', '', 'fv-ie-lora-note');
+    const add = el('button', t('Add', '添加')); add.type = 'button';
+    actions.append(add); panel.append(list, actions, note); menu.append(summary, panel);
+    summary.setAttribute('aria-label', 'LoRA');
+
+    const usable = () => activeLoras(rows, {names, verdicts});
+    function save() {
+        try { storage?.setItem(STORE, JSON.stringify({rows})); } catch { /* private mode */ }
+    }
+    function drawSummary() {
+        const count = usable().length;
+        summary.textContent = count ? `LoRA · ${count}` : 'LoRA';
+        menu.dataset.active = String(count > 0);
+    }
+    function syncPrompt() {
+        if (!setPrompt) return;
+        const next = reconcileLoraTriggers(getPrompt(), records, usable(),
+            Object.fromEntries([...words].filter(([, w]) => typeof w === 'string')));
+        records = next.records;
+        if (next.prompt !== getPrompt()) setPrompt(next.prompt);
+    }
+    function changed() { syncPrompt(); save(); drawSummary(); onChange(); }
+
+    async function triggerWords(name) {
+        if (words.has(name)) return words.get(name);
+        if (!loading.has(name)) loading.set(name, (async () => {
+            // The first read of a big file hashes it once on the server; keep asking while that finishes.
+            for (let attempt = 0; attempt < 10; attempt++) {
+                try {
+                    const response = await fetchApi('/freevideo/lora/triggers?' + new URLSearchParams({name}),
+                        {cache: 'no-store', signal: AbortSignal.timeout(6500)});
+                    const result = await response.json();
+                    if (response.ok && typeof result?.words === 'string') { words.set(name, result.words); return result.words; }
+                    break;
+                } catch (error) { if (error?.name !== 'TimeoutError') break; }
+            }
+            throw new Error(t('Could not read trigger words.', '无法读取触发词。'));
+        })().finally(() => loading.delete(name)));
+        return loading.get(name);
+    }
+    async function saveWords(name, value) {
+        let result = null;
+        try {
+            const response = await fetchApi('/freevideo/lora/triggers', {method: 'POST', headers: {'Content-Type': 'application/json'},
+                body: JSON.stringify({name, words: value.trim()})});
+            result = await response.json();
+            if (!response.ok || typeof result?.words !== 'string') result = null;
+        } catch { /* same sentence below */ }
+        if (!result) throw new Error(t('Could not save. Try again.', '没有保存，请再试一次。'));
+        words.set(name, result.words);
+    }
+
+    function problemText(name, verdict) {
+        if (names && !names.includes(name)) return t('The file is no longer in the LoRA folder.', '此文件已不在 LoRA 文件夹中。');
+        if (verdict?.usable !== false) return '';
+        return verdict.reason === 'unreadable'
+            ? t('The file could not be read. It may be incomplete; download it again.', '文件无法读取，可能未下载完整，请重新下载。')
+            : t('This LoRA cannot be used for pictures.', '此 LoRA 不适用于图片生成。');
+    }
+
+    function render() {
+        if (disposed) return;
+        list.replaceChildren();
+        rows.forEach((row, index) => {
+            const wrap = el('div', null, 'fv-ie-lora-row'); wrap.dataset.enabled = String(row.enabled !== false);
+            const on = el('input'); on.type = 'checkbox'; on.checked = row.enabled !== false;
+            on.setAttribute('aria-label', t('Enable LoRA', '启用 LoRA'));
+            const select = el('select'); select.setAttribute('aria-label', t('LoRA model', 'LoRA 模型'));
+            for (const name of [...new Set([...(names || []), row.name])]) { const option = el('option', name); option.value = name; select.append(option); }
+            select.value = row.name; select.title = row.name;
+            const strength = el('input'); strength.type = 'number'; strength.step = '.05';
+            strength.min = String(-IMAGE_LORA_STRENGTH); strength.max = String(IMAGE_LORA_STRENGTH); strength.value = row.strength;
+            strength.setAttribute('aria-label', t('LoRA strength', 'LoRA 强度'));
+            const remove = el('button', '×'); remove.type = 'button'; remove.setAttribute('aria-label', t('Remove LoRA', '移除 LoRA'));
+            const trigger = el('input', null, 'fv-ie-lora-trigger'); trigger.type = 'text'; trigger.maxLength = 300;
+            trigger.placeholder = t('Trigger words', '触发词'); trigger.setAttribute('aria-label', trigger.placeholder);
+            trigger.value = drafts.get(row.name) ?? words.get(row.name) ?? '';
+            const problem = el('p', '', 'fv-ie-lora-problem'); problem.hidden = true;
+            const show = () => {
+                const text = problemText(row.name, verdicts.peek(row.name));
+                problem.textContent = text; problem.hidden = !text;
+                wrap.dataset.unusable = String(!!text);
+                on.disabled = !!text; strength.disabled = trigger.disabled = !!text || row.enabled === false;
+            };
+            on.onchange = () => { row.enabled = on.checked; wrap.dataset.enabled = String(on.checked); show(); changed(); };
+            select.onchange = () => {
+                row.name = select.value; select.title = row.name; trigger.value = words.get(row.name) ?? '';
+                if (verdicts.peek(row.name)?.usable === false) row.enabled = true;
+                on.checked = row.enabled !== false; check(); changed();
+            };
+            strength.onchange = () => {
+                const value = Number(strength.value);
+                if (strength.value === '' || !(Math.abs(value) <= IMAGE_LORA_STRENGTH)) {
+                    strength.setCustomValidity(t(`Enter a value from -${IMAGE_LORA_STRENGTH} to ${IMAGE_LORA_STRENGTH}.`, `请输入 -${IMAGE_LORA_STRENGTH} 到 ${IMAGE_LORA_STRENGTH} 之间的数。`));
+                    strength.reportValidity(); return;
+                }
+                strength.setCustomValidity(''); row.strength = value; changed();
+            };
+            remove.onclick = () => { rows.splice(index, 1); render(); changed(); };
+            trigger.onchange = async () => {
+                const name = row.name, draft = trigger.value;
+                drafts.set(name, draft);
+                try { await saveWords(name, draft); drafts.delete(name); note.textContent = ''; syncPrompt(); save(); onChange(); }
+                catch (error) { note.textContent = error.message; }
+            };
+            const check = () => verdicts.get(row.name).then(() => { if (!disposed && wrap.isConnected && select.value === row.name) { show(); drawSummary(); syncPrompt(); onChange(); } });
+            wrap.append(on, select, strength, remove, trigger, problem);
+            list.append(wrap); show();
+            if (!verdicts.peek(row.name)) check();
+            triggerWords(row.name).then(() => {
+                if (!disposed && trigger.isConnected && !drafts.has(row.name)) trigger.value = words.get(row.name) ?? '';
+                syncPrompt(); save(); drawSummary(); onChange();
+            }, error => { note.textContent = error.message; });
+        });
+        add.disabled = rows.length >= IMAGE_LORA_LIMIT;
+        if (!note.textContent || !rows.some(r => drafts.has(r.name))) note.textContent = skipped
+            ? t(`${skipped} other LoRA file${skipped === 1 ? '' : 's'} in the folder ${skipped === 1 ? 'is' : 'are'} not in .safetensors format and not listed.`, `文件夹中另有 ${skipped} 个 LoRA 文件不是 .safetensors 格式，未列出。`) : '';
+        drawSummary();
+    }
+
+    async function refresh() {
+        try {
+            const response = await fetchApi('/models/loras', {cache: 'no-store'});
+            if (!response.ok) throw new Error(response.statusText);
+            const listed = (await response.json()).filter(n => typeof n === 'string');
+            skipped = listed.filter(n => /\.(ckpt|pt|pth|bin|sft)$/i.test(n)).length;
+            names = listed.filter(n => n.toLowerCase().endsWith('.safetensors'));
+            verdicts.clear(); render(); changed();
+        } catch (error) {
+            if (!disposed) note.textContent = t('Could not read the LoRA folder.', '无法读取 LoRA 文件夹。');
+        }
+    }
+    add.onclick = async () => {
+        await refresh();
+        if (rows.length >= IMAGE_LORA_LIMIT) { note.textContent = t(`Use at most ${IMAGE_LORA_LIMIT} LoRAs`, `最多使用 ${IMAGE_LORA_LIMIT} 个 LoRA`); return; }
+        if (!names?.length) { note.textContent = t('Put LoRA files (.safetensors) in the models/loras folder, then choose "Add".', '把 LoRA 文件（.safetensors）放进 models/loras 文件夹，再点“添加”。'); return; }
+        rows.push({name: names.find(n => !rows.some(r => r.name === n)) ?? names[0], strength: 1, enabled: true});
+        render(); changed();
+    };
+    menu.addEventListener('toggle', () => { if (menu.open) refresh(); });
+    render();
+    return {
+        element: menu,
+        /** The LoRAs to send with the request; empty when none is switched on. */
+        value: usable,
+        count: () => usable().length,
+        refresh,
+        /** After the prompt box is cleared or replaced, trigger words are put back for the rows still on. */
+        reconcile: syncPrompt,
+        dispose() { disposed = true; },
+    };
+}
diff --git a/web/lora_check.js b/web/lora_check.js
index 1889d7a4..344c1f07 100644
--- a/web/lora_check.js
+++ b/web/lora_check.js
@@ -43,7 +43,7 @@ export function feedbackDraft(name, verdict, described, t = (en) => en) {
 }
 
 /** Verdicts by name. Refresh also checks the currently installed model; failures say nothing. */
-export function createVerdicts(fetchApi) {
+export function createVerdicts(fetchApi, {kind = null} = {}) {
     let known = new Map(), pending = new Map();
     return {
         peek: name => known.get(name),
@@ -53,7 +53,7 @@ export function createVerdicts(fetchApi) {
             if (!pending.has(name)) {
                 const mine = pending, request = (async () => {
                     try {
-                        const response = await fetchApi('/freevideo/lora/check?' + new URLSearchParams({name}), {cache: 'no-store'});
+                        const response = await fetchApi('/freevideo/lora/check?' + new URLSearchParams(kind ? {name, kind} : {name}), {cache: 'no-store'});
                         const verdict = response.ok ? await response.json() : null;
                         if (mine !== pending) return null;
                         if (verdict && typeof verdict.usable === 'boolean') known.set(name, verdict);
diff --git a/web/tutorial/replace-person-1.webp b/web/tutorial/replace-person-1.webp
new file mode 100644
index 00000000..aa5c3307
Binary files /dev/null and b/web/tutorial/replace-person-1.webp differ
diff --git a/web/tutorial/replace-person-2.webp b/web/tutorial/replace-person-2.webp
new file mode 100644
index 00000000..4c8c71be
Binary files /dev/null and b/web/tutorial/replace-person-2.webp differ
diff --git a/web/tutorial_book_content.js b/web/tutorial_book_content.js
index 8f823a76..c4090870 100644
--- a/web/tutorial_book_content.js
+++ b/web/tutorial_book_content.js
@@ -8,10 +8,10 @@ export const TUTORIAL_CHAPTERS = [
         media: {
             type: 'image',
             src: {zh: 'quick-start-1.webp', en: 'quick-start-2.webp'},
-            alt: {zh: '创作面板：描述框和“生成视频”按钮已用红框标出。', en: 'The creation panel, with the prompt box and the "Generate video" button outlined in red.'},
+            alt: {zh: '创作面板：描述框和“生⁠成⁠视⁠频”按钮已用红框标出。', en: 'The creation panel, with the prompt box and the "Generate video" button outlined in red.'},
         },
         steps: {
-            zh: ['打开“创作面板”，在“描述画面”中写一句话，说明谁在什么地方做什么。', '画幅选“横屏”或“竖屏”，其余保持默认。', '点“生成视频”。', '完成后在预览区播放，点“下载视频”保存。'],
+            zh: ['打开“创作面板”，在“描述画面”中写一句话，说明谁在什么地方做什么。', '画幅选“横屏”或“竖屏”，其余保持默认。', '点“生⁠成⁠视⁠频”。', '完成后在预览区播放，点“下载视频”保存。'],
             en: ['Open "Create" and write one sentence under the prompt: who is where and doing what.', 'Choose "Landscape" or "Portrait" and keep the rest as it is.', 'Choose "Generate video".', 'When it finishes, play it in the preview and choose "Download video" to save it.'],
         },
         example: {
@@ -20,8 +20,8 @@ export const TUTORIAL_CHAPTERS = [
             settings: {},
         },
         limits: {
-            zh: ['首次生成需要加载模型，用时较长；之后会更快。', '同一时间只处理一个请求，其余请求在队列中等待。'],
-            en: ['The first run loads the models and takes longer. Later runs are faster.', 'One request runs at a time and the others wait in the queue.'],
+            zh: ['首次生成需要加载模型，用时较长；之后会更快。', '每张显卡同一时间处理一个请求，其余请求在队列中等待。'],
+            en: ['The first run loads the models and takes longer. Later runs are faster.', 'Each graphics card runs one request at a time, and the others wait in the queue.'],
         },
     },
     {
@@ -34,8 +34,8 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '创作面板：描述、画幅与时长区域已用红框标出。', en: 'The creation panel, with the prompt and the Frame & duration fields outlined in red.'},
         },
         steps: {
-            zh: ['在“描述画面”中写明主体、动作、场景和镜头，例如“镜头缓缓推近”。', '在“画幅与时长”中选择横屏、竖屏或方形，并设置“清晰度”。', '在“时长 · 秒”中填写秒数；“可生成范围”中可查看各尺寸的最长时长和预计用时。', '点“生成视频”。'],
-            en: ['Write the subject, action, place and camera in the prompt, for example "the camera slowly moves closer".', 'Under "Frame & duration", choose landscape, portrait or square and set "Resolution".', 'Enter the seconds under "Duration · seconds". "Available sizes and lengths" shows the longest length and the estimated time for each size.', 'Choose "Generate video".'],
+            zh: ['在“描述画面”中写明主体、动作、场景和镜头，例如“镜头缓缓推近”。', '在“画幅与时长”中选择横屏、竖屏或方形，并设置“清晰度”。', '在“时长 · 秒”中填写秒数；“可⁠生⁠成⁠范⁠围”中可查看各尺寸的最长时长和预计用时。', '点“生⁠成⁠视⁠频”。描述为空时会先提示，点“继⁠续⁠生⁠成”仍可生成。'],
+            en: ['Write the subject, action, place and camera in the prompt, for example "the camera slowly moves closer".', 'Under "Frame & duration", choose landscape, portrait or square and set "Resolution".', 'Enter the seconds under "Duration · seconds". "Available sizes and lengths" shows the longest length and the estimated time for each size.', 'Choose "Generate video". With an empty prompt, a note appears first; choose "Generate anyway" to go ahead.'],
         },
         example: {
             assets: [],
@@ -57,7 +57,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '创作面板：素材区域中图片的用途下拉框已用红框标出。', en: 'The creation panel, with the image role selectors in the Media section outlined in red.'},
         },
         steps: {
-            zh: ['在“素材”中点“上传文件”，添加起始图片，并将它的用途下拉框选为“首帧”。', '需要结尾画面时，再添加一张图片并设为“尾帧”。', '在描述中写明两张图之间发生了什么。', '点“生成视频”。'],
+            zh: ['在“素材”中点“上传文件”，添加起始图片，并将它的用途下拉框选为“首帧”。', '需要结尾画面时，再添加一张图片并设为“尾⁠帧”。', '在描述中写明两张图之间发生了什么。', '点“生⁠成⁠视⁠频”。'],
             en: ['Under "Media", choose "Upload files", add the starting image and set its role dropdown to "First frame".', 'For an ending picture, add another image and set it to "Last frame".', 'Describe what happens between the two images.', 'Choose "Generate video".'],
         },
         example: {
@@ -80,8 +80,8 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '创作面板：素材区域和参考图已用红框标出。', en: 'The creation panel, with the Media section and a reference image outlined in red.'},
         },
         steps: {
-            zh: ['在创作面板的“素材”中添加图片，每张可设为参考、首帧或尾帧。', '在描述中输入 @ 选择图片，并写明它在画面中的作用，例如“<Picture 1> 里的女孩在海边奔跑”。', '视频和音频也可作为参考：点“上\u2060传\u2060文\u2060件”添加，在描述中输入 @ 引用。', '点“生成视频”。'],
-            en: ['Add images under "Media" in the creation panel and set each one as a reference, first frame, or last frame.', 'Type @ in the prompt and choose an image, then say what it does in the scene, for example "The girl in <Picture 1> runs along the beach".', 'Videos and audio can be references too: add them with "Upload files" and mention them in the prompt with @.', 'Choose "Generate video".'],
+            zh: ['在创作面板的“素材”中添加图片，每张可设为参考、首帧或尾帧。', '在描述中输入 @ 选择图片，并写明它在画面中的作用，例如“<Picture 1> 里的女孩在海边奔跑”。', '视频和音频也可作为参考：点“上\u2060传\u2060文\u2060件”添加，在描述中输入 @ 引用。', '音频卡片上选“原样使用”时，成片直接使用这段声音，时长随音频而定；一次只能有一段音频原样使用。', '点“生⁠成⁠视⁠频”。'],
+            en: ['Add images under "Media" in the creation panel and set each one as a reference, first frame, or last frame.', 'Type @ in the prompt and choose an image, then say what it does in the scene, for example "The girl in <Picture 1> runs along the beach".', 'Videos and audio can be references too: add them with "Upload files" and mention them in the prompt with @.', 'On an audio card, choose "Use as is" to put that audio into the result unchanged. The duration follows the audio, and only one audio file can be used as is.', 'Choose "Generate video".'],
         },
         example: {
             assets: [],
@@ -103,7 +103,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '角色窗口：名字、正面图片和其他角度图片已用红框标出。', en: 'The character window, with the name, front image and other angles outlined in red.'},
         },
         steps: {
-            zh: ['展开创作面板的“素材”，在“角色”一行点“+”（新建角色）。', '填写“名字”，设置“正面图片”：可“上传图片”，也可“AI 生成”，后者会按描述一次生成 5 张供选。', '可再添加最多 5 张其他角度的图片，然后点“保存”。', '之后在描述中输入 @ 选择该角色，并写明其动作。'],
+            zh: ['展开创作面板的“素材”，在“角色”一行点“+”（新建角色）。', '填写“名字”，设置“正面图片”：可“上⁠传⁠图⁠片”，也可“AI 生成”，后者会按描述一次生成 5 张供选。', '可再添加最多 5 张其他角度的图片，然后点“保⁠存”。', '之后在描述中输入 @ 选择该角色，并写明其动作。'],
             en: ['Expand "Media" in the creation panel and, on the "Characters" row, choose + (New character).', 'Enter a "Name" and set the "Front image": choose "Upload images", or "Generate with AI", which makes 5 pictures from a description to pick from.', 'Add up to 5 images from other angles, then choose "Save".', 'In later prompts, type @ to pick the character and describe what the character does.'],
         },
         example: {
@@ -116,6 +116,24 @@ export const TUTORIAL_CHAPTERS = [
             en: ['A request uses up to 3 characters.', 'Characters cannot be used with first or last frames.', 'A character may look slightly different between runs.'],
         },
     },
+    {
+        id: 'replace-person',
+        title: {zh: '替换人物', en: 'Replace person'},
+        summary: {zh: '把视频里的人换成人物图中的人，动作和镜头沿用源视频。', en: 'Put the person from an image into a video, keeping the motion and camera of the source video.'},
+        media: {
+            type: 'image',
+            src: {zh: 'replace-person-1.webp', en: 'replace-person-2.webp'},
+            alt: {zh: '创作面板：素材中的“替换人物”、源视频和人物图两格、“保⁠留⁠源⁠视⁠频⁠声⁠音”已用红框标出。', en: 'The creation panel, with "Replace person", the source video and person image slots, and "Keep source video sound" outlined in red.'},
+        },
+        steps: {
+            zh: ['展开创作面板的“素材”，打开“替换人物”。', '在“源视频”中点“添加视频”，在“人物图”中点“添加图片”。', '需要源视频的声音时，勾选“保⁠留⁠源⁠视⁠频⁠声⁠音”。', '描述可以留空，也可以补充画面要求；然后点“生⁠成⁠视⁠频”。'],
+            en: ['Expand "Media" in the creation panel and turn on "Replace person".', 'Under "Source video" choose "Add video", and under "Person image" choose "Add image".', 'To keep the sound of the source video, check "Keep source video sound".', 'The prompt can stay empty or describe more of the picture. Then choose "Generate video".'],
+        },
+        limits: {
+            zh: ['替换人物不能与首帧、尾帧同时使用。', '首次使用时会自动下载替换人物模型（约 80 MB），面板中显示进度，下载完成后才能生成。', '源视频至少 1 秒；源视频没有声音时，不能保留声音。'],
+            en: ['Replace person cannot be used with first or last frames.', 'The first time, the Replace person model (about 80 MB) downloads on its own. The panel shows the progress, and generation can start when it finishes.', 'The source video must be at least 1 second long. A source video without sound has no sound to keep.'],
+        },
+    },
     {
         id: 'quality-tiers',
         title: {zh: '四个质量档位', en: 'The four quality levels'},
@@ -126,7 +144,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '创作面板：质量滑块上的轻量、标准、精细、极致已用红框标出。', en: 'The creation panel, with the quality slider and its four names outlined in red.'},
         },
         steps: {
-            zh: ['在“质量”滑块上选择“轻量”“标准”“精细”或“极致”。', '构思阶段用“轻量”，快速确认画面和动作。', '定稿时改用“精细”或“极致”。', '选好档位后，“质量”旁会显示预计用时；然后点“生成视频”。'],
+            zh: ['在“质量”滑块上选择“轻量”“标准”“精细”或“极⁠致”。', '构思阶段用“轻量”，快速确认画面和动作。', '定稿时改用“精细”或“极⁠致”。', '选好档位后，“质量”旁会显示预计用时；然后点“生⁠成⁠视⁠频”。'],
             en: ['Move the "Quality" slider to Light, Medium, High or Max.', 'While exploring, use Light to check the picture and motion quickly.', 'For the final video, switch to High or Max.', 'The estimated time for the chosen level appears next to "Quality". Then choose "Generate video".'],
         },
         example: {
@@ -149,7 +167,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '创作面板：二次采样选项“先出预览”已用红框标出。', en: 'The creation panel, with the "Preview first" option outlined in red.'},
         },
         steps: {
-            zh: ['在二次采样中选择“先出预览”，按钮变为“生成预览”。', '点“生成预览”，等待预览完成。', '画面满意时，点结果中的“继续二采”。', '不满意时，修改描述或种子，再生成新的预览。'],
+            zh: ['在二次采样中选择“先出预览”，按钮变为“生⁠成⁠预⁠览”。', '点“生⁠成⁠预⁠览”，等待预览完成。', '画面满意时，点结果中的“继续二采”。', '不满意时，修改描述或种子，再生成新的预览。'],
             en: ['Under two-pass sampling choose "Preview first". The button changes to "Generate preview".', 'Choose "Generate preview" and wait for it to finish.', 'When the picture is right, choose "Run second pass" in the result.', 'Otherwise change the prompt or the seed and generate a new preview.'],
         },
         example: {
@@ -169,10 +187,10 @@ export const TUTORIAL_CHAPTERS = [
         media: {
             type: 'image',
             src: {zh: 'long-video-1.webp', en: 'long-video-2.webp'},
-            alt: {zh: '创作面板：时长输入框和“可生成范围”入口已用红框标出。', en: 'The creation panel, with the duration field and "Available sizes and lengths" control outlined in red.'},
+            alt: {zh: '创作面板：时长输入框和“可⁠生⁠成⁠范⁠围”入口已用红框标出。', en: 'The creation panel, with the duration field and "Available sizes and lengths" control outlined in red.'},
         },
         steps: {
-            zh: ['填写更长的时长，会自动分段生成；“可生成范围”中可查看各尺寸的最长时长和预计用时。', '点“生成视频”。', '在生成好的视频下方点“接着生成”，继续下一段。', '更长的多镜头视频，可在“剪辑视图”中另起新的镜头，或使用“AI 全自动”。'],
+            zh: ['填写更长的时长，会自动分段生成；“可⁠生⁠成⁠范⁠围”中可查看各尺寸的最长时长和预计用时。', '点“生⁠成⁠视⁠频”。', '在生成好的视频下方点“接着生成”，继续下一段。', '更长的多镜头视频，可在“剪辑视图”中另起新的镜头，或使用“AI 全自动”。'],
             en: ['Enter a longer duration. It is made in parts automatically. "Available sizes and lengths" shows how long each size can be and about how long it takes.', 'Choose "Generate video".', 'Choose "Extend" below a finished video to continue it.', 'For a longer video with several shots, start new shots in "Edit", or use "Autopilot".'],
         },
         example: {
@@ -195,7 +213,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '剪辑视图：接着生成的草稿段和描述框已用红框标出。', en: 'The Edit view, with the draft part that continues the video and the prompt box outlined in red.'},
         },
         steps: {
-            zh: ['在生成好的视频下方点“接着生成”，打开剪辑视图。', '选中后面自动加上的那一段（默认 10 秒；在 Mac 和一次生成不了 10 秒的设备上为 5 秒），在底部描述框写明接下来发生什么。', '需要更长时，拖长这一段，每段至少 5 秒，然后点“生成”。', '新的一段会接在原视频之后，并自动拼成一条。'],
+            zh: ['在生成好的视频下方点“接着生成”，打开剪辑视图。', '选中后面自动加上的那一段（默认 10 秒；在 Mac 和一次生成不了 10 秒的设备上为 5 秒），在底部描述框写明接下来发生什么。', '需要更长时，拖长这一段，每段至少 5 秒，然后点“生⁠成”。', '新的一段会接在原视频之后，并自动拼成一条。'],
             en: ['Choose "Extend" below a finished video. The Edit view opens.', 'Select the part added after it (10 seconds by default, or 5 seconds on a Mac and on devices that cannot make 10 seconds at once) and write what happens next in the box at the bottom.', 'To make it longer, drag the part out; each part is at least 5 seconds. Then choose "Generate".', 'The new part follows the original and is joined into one video.'],
         },
         example: {
@@ -218,7 +236,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '剪辑视图时间线：接续上一段的草稿带有连接标记，与后面空开的独立镜头已用红框标出。', en: 'The Edit timeline, with the draft that continues the video, marked by a link, and the separate shot after a gap outlined in red.'},
         },
         steps: {
-            zh: ['要让画面接着上一段往下发展，在生成好的视频下方点“接着生成”：新的一段紧接在后面，带有连接标记，与前面拼成同一个镜头。', '要换场景或切到另一个画面，在时间线的空白处拖出一段，写下描述再点“生成”：它是独立镜头，不接前面的画面。', '看时间线：接续的一段紧贴前一段，带连接标记；独立镜头与前面的视频分开显示。'],
+            zh: ['要让画面接着上一段往下发展，在生成好的视频下方点“接着生成”：新的一段紧接在后面，带有连接标记，与前面拼成同一个镜头。', '要换场景或切到另一个画面，在时间线的空白处拖出一段，写下描述再点“生⁠成”：它是独立镜头，不接前面的画面。', '看时间线：接续的一段紧贴前一段，带连接标记；独立镜头与前面的视频分开显示。'],
             en: ['To carry the picture on from the previous part, choose "Extend" below a finished video. The new part follows right after it, carries a link mark, and joins the earlier part as one shot.', 'To change scene or cut to another picture, drag out a part in an empty place on the timeline, describe it and choose "Generate". It is a separate shot and does not continue the earlier picture.', 'Check the timeline: a continuing part sits right against the part before it and carries a link mark. A separate shot is drawn apart from the video before it.'],
         },
         example: {
@@ -241,7 +259,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '剪辑视图：选中的一段和描述框已用红框标出。', en: 'The Edit view, with the selected part and the prompt box outlined in red.'},
         },
         steps: {
-            zh: ['点顶部的“剪辑视图”。', '在时间线上选中要处理的一段。', '在下方描述框中填写这一段的内容，点“生成”接续下一段。', '在右侧“画面”中可调整缩放、位置和旋转，或点“裁剪”（窗口窄于 1280 时，先点“调整”打开该面板）。', '完成后点右上角的“导出”保存。'],
+            zh: ['点顶部的“剪辑视图”。', '在时间线上选中要处理的一段。', '在下方描述框中填写这一段的内容，点“生⁠成”接续下一段。', '在右侧“画面”中可调整缩放、位置和旋转，或点“裁⁠剪”（窗口窄于 1280 时，先点“调⁠整”打开该面板）。', '完成后点右上角的“导⁠出”保存。'],
             en: ['Choose "Edit" at the top.', 'Select a part on the timeline.', 'Write what happens in this part in the box below and choose "Generate" to continue.', 'Under "Video" on the right, adjust scale, position and rotation, or choose "Crop" (on a window narrower than 1280 pixels, choose "Adjust" first to open the panel).', 'When done, choose "Export" at the top right.'],
         },
         example: {
@@ -264,7 +282,7 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '剪辑视图：选中的接续段和描述框旁的参考图入口已用红框标出。', en: 'The Edit view, with the selected continuation and the reference image control beside the prompt outlined in red.'},
         },
         steps: {
-            zh: ['选中要接续的一段。', '点描述框旁的 +，上传或生成图片，最多 3 张。', '填写描述，然后点“生成”。'],
+            zh: ['选中要接续的一段。', '点描述框旁的 +，上传或生成图片，最多 3 张。', '填写描述，然后点“生⁠成”。'],
             en: ['Select the part to continue.', 'Click + next to the prompt to upload or generate images, up to 3.', 'Describe it, then choose "Generate".'],
         },
         example: {
@@ -287,12 +305,12 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: 'AI 全自动：连接 Agent 的按钮已用红框标出。', en: 'Autopilot, with the button that connects an agent outlined in red.'},
         },
         steps: {
-            zh: ['点顶部的“AI 全自动”。', '点“连接 Agent”卡片中的按钮复制连接信息，粘贴给所使用的 Agent。', '状态显示“已连接”（或所连 Agent 的名字）后，在“对 Agent 说…”中描述想做的视频，例如“做一条 1 分钟的竖屏短片：咖啡店的一天”。', '点“发送”，Agent 会生成并剪辑；可随时点“停止”，完成后点“打开”或“剪辑”。'],
-            en: ['Choose "Autopilot" at the top.', 'On the "Connect an agent" card, choose the button to copy the connection details and paste them to the agent you use.', 'When it shows "Connected" (or the agent\'s name), describe the video under "Message the agent…", for example "Make a 1-minute vertical short film: a day in a coffee shop".', 'Choose "Send". The agent generates and edits it. Choose "Stop" at any time, and "Open" or "Edit" when it finishes.'],
+            zh: ['点顶部的“AI 全自动”。', '点“连接 Agent”卡片中的按钮复制连接信息，粘贴给所使用的 Agent。', '状态显示“已连接”（或所连 Agent 的名字）后，在“对 Agent 说…”中描述想做的视频，例如“做一条 1 分钟的竖屏短片：咖啡店的一天”。', '点“发送”，Agent 会生成并剪辑；可随时点“停止”，完成后点“打开”或“剪辑”。', '在对话列表中点对话旁的“更多”，选“重命名”可修改对话名称。'],
+            en: ['Choose "Autopilot" at the top.', 'On the "Connect an agent" card, choose the button to copy the connection details and paste them to the agent you use.', 'When it shows "Connected" (or the agent\'s name), describe the video under "Message the agent…", for example "Make a 1-minute vertical short film: a day in a coffee shop".', 'Choose "Send". The agent generates and edits it. Choose "Stop" at any time, and "Open" or "Edit" when it finishes.', 'In the chat list, choose "More" next to a chat, then "Rename" to change its name.'],
         },
         limits: {
-            zh: ['需要先连接 Agent，未连接时只显示连接说明。', '镜头越多，用时越长。'],
-            en: ['An agent must be connected first. Until then only the connection card is shown.', 'More shots take longer.'],
+            zh: ['需要先连接 Agent，未连接时只显示连接说明。', '镜头越多，用时越长。', 'Agent 开拍整部片前、重拍耗时较长时是否先询问，可在“设置”的“AI Agent”中选择。'],
+            en: ['An agent must be connected first. Until then only the connection card is shown.', 'More shots take longer.', 'Whether an agent asks before starting a film or a long reshoot is set under "AI Agent" in Settings.'],
         },
     },
     {
@@ -305,8 +323,8 @@ export const TUTORIAL_CHAPTERS = [
             alt: {zh: '“我的作品”：第一条作品已用红框标出。', en: 'Your creations, with the first video outlined in red.'},
         },
         steps: {
-            zh: ['点预览区的“作品”，打开“我的作品”。', '选择一条视频，可播放，或点“接着生成”在剪辑视图中继续。', '需要成片时，在“剪辑视图”点“导出”。', '在导出窗口选择范围、画质和内容，再点“导出”。'],
-            en: ['Choose "Creations" in the preview area to open the library.', 'Pick a video to play it, or choose "Extend" to continue it in the Edit view.', 'For a finished file, choose "Export" in the Edit view.', 'In the export window choose the range, quality and content, then choose "Export".'],
+            zh: ['点预览区的“作品”，打开“我的作品”。', '选择一条视频，可播放，或点“接着生成”在剪辑视图中继续。', '在作品详情中点爱心图标收藏；列表上方选“收⁠藏”只看收藏的作品，选“全⁠部”看所有作品。', '需要成片时，在“剪辑视图”点“导⁠出”。', '在导出窗口选择范围、画质和内容，再点“导⁠出”。'],
+            en: ['Choose "Creations" in the preview area to open the library.', 'Pick a video to play it, or choose "Extend" to continue it in the Edit view.', 'In a creation\'s details, choose the heart to add it to favorites. Above the list, choose "Favorites" to see only those, or "All" to see every creation.', 'For a finished file, choose "Export" in the Edit view.', 'In the export window choose the range, quality and content, then choose "Export".'],
         },
         limits: {
             zh: ['删除作品前会确认；被剪辑使用的作品不能删除。', '导出过程中取消不会留下不完整的文件。'],
@@ -327,8 +345,8 @@ export const TUTORIAL_CHAPTERS = [
             en: ['Read the error message first. It names what to try.', 'If it still fails, choose "Export report" to save the diagnostic information.', 'Choose "Report a problem" and attach the report to a GitHub issue.', 'When asked to restart, restart FreeVideo.'],
         },
         limits: {
-            zh: ['报告中的路径已做脱敏处理。'],
-            en: ['Paths in the report are anonymized.'],
+            zh: ['报告中的路径已做脱敏处理。', '全部下载源都失败时，错误卡片会逐个列出下载源和失败原因。'],
+            en: ['Paths in the report are anonymized.', 'When every download source fails, the error card lists each source and the reason it failed.'],
         },
     },
 ];