Carry the details of a reference into your next image.
Jingzao Image Forge (镜造 Image Forge) turns an image brief and selected reference evidence into an inspectable visual specification, platform-ready prompt, and review criteria. It helps keep the decisions that often drift—hair colour, expression, gaze, action, visible body design, wardrobe, camera, light, material, and required text—explicit and reviewable.
简体中文 · Skill instructions · Practical guide · AI-readable project map · Project-presentation research
New introduction artwork; actual generated samples and facet-level reviews appear below.
Transfer lilac-white studio depth, controlled dark-shell reflections, distinct dry-pigment/fiber materials and product-led element effects to new products. Artist pastels, an electronics array and a contact macro were tested. The refined pastel still has some coarse particles; the record retains initial failures and actual canvas sizes.
| Electronics without powder effects | Dry-pigment contact macro |
|---|---|
 |
 |
Style guide and research method · Refined pastel sample · Five generation/refinement records
Install into the user-level Codex Skills directory:
git clone https://github.com/papperrollinggery/jingzao-image-forge.git ~/.codex/skills/jingzao-image-forge
cd ~/.codex/skills/jingzao-image-forge
python3 scripts/verify_install.py . . --ref HEADmatch from verify_install.py checks the tracked files and symlinks against the resolved Git commit. Start a new Codex task before checking Skill discovery; an install check does not show that an already-open task adopted the Skill.
For a natural request, call the Skill directly:
$jingzao-image-forge Create a 21:9 film-frame key visual from this brief.
Return a visual_generation_spec, an OpenAI-ready prompt, and the review points.
For an existing specification, compile an OpenAI prompt:
python3 scripts/compile_prompt.py tests/forward-specs/minimal-product-neutral.json --platform openaiThe v1.10 reference-profile route adds selected visual evidence before compilation:
python3 scripts/compile_prompt.py SPEC.json \
--reference-profile PROFILE.json \
--platform openai--reference-profile and --style-capsule are mutually exclusive. The former projects selected, observable reference evidence into the target; the latter applies reusable style rules while keeping the target specification authoritative. See the guide for the contract and the prompt compiler reference for the provider output.
- A maintainable
visual_generation_specfor a new image, edit, restyle, expansion, cinematic frame, styleboard, or production image requirement. - A reference profile that separates what is observed, unknown, preserved, adapted, omitted, or not applicable across 14 facets:
face_design,hair,makeup,expression,gaze,pose,action,body,wardrobe,camera,lighting,palette,surface, andenvironment. - Platform-ready prompt packages for OpenAI, FLUX, Midjourney, or a generic output, without inventing provider controls.
- Reusable, source-image-free style capsules with explicit transfer boundaries.
- Visible-state and camera requirements for film or video pre-production image coverage.
Jingzao is designed for practical still-image work: character and wardrobe studies, product and material images, architecture, key art, narrative frames, action, interface-state boards, and reference-aware edits. It scales the structure to the brief; a simple request does not need a full production package.
| Mode | Use it for |
|---|---|
create |
A new image from a brief. |
reconstruct |
Observable relationships from an actual source image. |
edit |
A named local change with explicit preservation. |
restyle |
A new treatment while source content stays authoritative. |
expand |
New canvas edges with source-bound continuity checks. |
learn_style |
Reusable visual rules with transfer boundaries. |
styleboard |
Related cinematic frames or interface states. |
A reference image can carry different responsibilities. A reference profile records the responsibility before it becomes prompt prose. For each of the 14 facets, it captures evidence, a decision, an optional positive projection, and a review check. unknown stays unknown; it is not promoted into a preserved fact.
For attached profiles, the named source must also be a real must_attach: true input in the target specification. The compiled package can describe a handoff plan, but a description is not a delivered image. For text_transfer, the profile must not claim an attachment was sent.
This keeps three things separate:
- Observed evidence — what a reviewer can actually see in the source.
- Creative selection — what the target should preserve, adapt, omit, or leave unresolved.
- Execution and acceptance — whether a real input was sent, and whether the returned image meets the intended visual checks.
Public v1.10 refinement sample. This text-only profile trial returned an image and improved fingertip-to-cheek contact and plain hanging ribbons relative to its first text-only attempt. It remains a partial result: the fringe is looser and grayer than the source, the headwear reads as lace rather than folded jacquard, and its top is cropped. Open the full-size image · Read the visual-review evidence
Two attached-reference close portraits also retained their main reference traits. An additional window-side body/hand-interaction call was input-blocked without an image. The successful calls do not replay the previously blocked inputs unchanged.
When a still-image package supports a video project, Jingzao can carry only the image-side handoff: a visible state, camera proof, reference roles, prompt, and review points. The surrounding method is:
video evidence → mechanism and controls → still-image specification → visual review
The first stage can use video-evidence-workbench; the mechanism/control stage can use film-breakdown-distiller; Jingzao owns the still-image portion. Those external Skills are optional. The reusable guidance lives in this repository's production coverage, shot design, and quality controls.
- Jingzao guides the image tools available in the current Codex host to generate and review images. Its Python scripts compile specifications and prompts; they are not an independently hosted rendering service.
- A valid JSON file, a successful command, a declared attachment, or a provider call plan is not visual acceptance.
- The project does not promise that a prompt will pass a provider's policy checks, reproduce an identity, preserve pixels, or produce the same result twice.
- Provider behavior, current controls, and model availability can change. Recheck a provider's current documentation before relying on a provider-specific route.
The concise entry point deliberately links outward instead of hiding the technical material.
| Need | Read |
|---|---|
| Reference-profile contract, current cases, and historical asset index | Practical guide · Reference distillation |
| Specification fields | Visual specification |
| Prompt compilation and provider boundaries | Prompt compiler |
| Reconstruct, restyle, and expand | Reference modes |
| Style learning and capsules | Style learning · Style capsules |
| Film stills and coverage | Production coverage · Cinematic shot design |
| Clean rendering and visual review | Quality controls · Image diagnostics |
| Existing examples and retained evidence | examples/ · tests/ |
Release checks: 255 automated tests, the complete local CI workflow, four-platform reference-trait projection, and independent code/visual review. Automated checks cover structure, compatibility and evidence binding; image quality is reviewed separately.
Earlier introduction graphics and gallery cases remain available from the historical-material index. They are isolated from the v1.10 entry image because they document earlier workflows and evidence conditions. They do not stand in for a v1.10 generation result.