VRChat DLSS5 Cam

English · 简体中文 · 日本語 · 한국어

New here? The documentation site walks you from the download to your first picture step by step, with pictures, in English, 简体中文, 日本語 and 한국어. The Documentation button in the application opens it in the interface's language, and Ask AI beside it answers questions from it without leaving the application (see AI Q&A). This page is the full reference.

VRChat DLSS5 Cam applies NVIDIA DLSS 5 Neural Rendering to the output of VRChat's camera on a GeForce RTX card and records the result as a lossless PNG. Images and videos on disk pass through the same pipeline, individually or as a batch. It is an ordinary Windows application: the VRChat process is never touched, and no mod is required.

Features

• Live camera. VRChat's Stream Camera is received over Spout and previewed in real time with DLSS 5 applied. A global hotkey (Ctrl+Alt+P) saves a lossless PNG while VRChat stays in the foreground.
• Images and video files. Screenshots and recordings can be dropped onto the window, adjusted and exported: PNG for stills, MP4 (H.264 / HEVC, audio track preserved), a PNG sequence or an animated GIF / APNG / WebP for video, named by a template of your own. Animated GIF, APNG and WebP files are processed frame by frame like videos and come back in the same format, with the timing of every frame kept.
• Batch processing. Every file that is opened is collected in the library below the preview. Any selection can be processed in a single run, either with the shared parameters or with per-file parameters, and each file can be turned, mirrored or cropped beforehand.
• Real DLSS 5 guidance. Motion vectors from NVIDIA Optical Flow, or from the FSR optical flow of this program on any card, are supplied to the network, so video and the live camera are processed with the same temporal cues a game would provide. An optional DLAA pass can clean up edges first; it and the depth view get a depth map from Depth Anything V2 (DLSS 5 itself reads only the picture and the motion vectors).
• Comparison and inspection. Wipe, original, depth and motion-vector views, wheel zoom, drag panning and a fullscreen mode make the difference before and after processing verifiable at 1:1.
• Built for daily use. A start page that leads through the first steps, a flat interface in dark and light themes that follow the Windows setting, one Advanced switch for everything beyond the essentials, undo and redo with a full history, user-defined presets, a search across every setting, a resizable layout that stays readable in a narrow sidebar, an automatic update check, questions about the program answered from its documentation in a side panel (Ask AI), four languages (English, 简体中文, 日本語, 한국어), a dedicated interface thread for a consistently responsive window, and a command line for scripted runs.
• Light on the computer. While nothing changes, the window redraws only a few times a second; the live camera is read only when VRChat has sent a new frame; and the depth network is loaded only while DLAA, super resolution or the depth view needs it.

Requirements

Windows	Windows 10 21H2 or Windows 11, 64-bit
Graphics card	NVIDIA GeForce RTX. The RTX 50 series and the RTX 40 / 30 / 20 series each have a build of the runtime in the archive (next row), so nothing has to be added for any of them. AMD Radeon RX 7000 / 9000 with the Radeon edition (VRChatDLSS5Cam-win64-amd.zip), which runs the neural pass through DLSS-NR-on-AMD, a separate project installed from the application (see AMD Radeon cards). Cards from other vendors are not refused: the application works as a viewer and recorder on them. DLAA and the hardware optical flow stay NVIDIA-only; the FSR optical flow takes over as the motion source.
DLSS 5 runtime	Included. The archive carries nvngx_dlssnr.dll 310.8.0.0 in two builds: runtimes\blackwell\ holds the build as shipped with games (RTX 50) and runtimes\universal\ a community-adapted build of the same runtime for RTX 40 / 30 / 20. The build for the installed card is chosen at start and the other is tried when it fails. Both files are NVIDIA's software under NVIDIA's terms and are not part of this project's MIT-licensed source (see THIRD_PARTY_NOTICES.md); nothing has to be obtained from anywhere else.
VRChat	Any build with the Stream Camera Spout Stream option (desktop or VR). Required for the live camera only.
AI Q&A (optional)	An internet connection and the Microsoft Edge WebView2 Runtime, which Windows 11 and most Windows 10 installations already have. Without the runtime, Ask AI opens the documentation site's AI Q&A in the browser instead.
Video files	Windows Media Foundation (part of Windows). The N / KN editions require the Media Feature Pack; HEVC files may require the HEVC Video Extensions from the Microsoft Store. Animated GIF, APNG and WebP files need nothing extra: the application decodes and writes them itself.

Getting started

Your graphics card	Download	Notes
NVIDIA GeForce RTX 20 / 30 / 40 / 50	VRChatDLSS5Cam-win64.zip	The DLSS 5 runtime is in the archive.
AMD Radeon RX 7000 / 9000	VRChatDLSS5Cam-win64-amd.zip	The Radeon edition; it fetches DLSS-NR-on-AMD's installer for you at the first start. See AMD Radeon cards for what to expect.
Linux with an NVIDIA card	VRChatDLSS5Cam-linux-x86_64.tar.gz	The GeForce edition running under Proton, with a launcher that sets Proton up by itself. Pictures, videos and MCP work; there is no live (Spout) mode and videos are saved as WebP, GIF, APNG or PNG. See Linux.

1. Download VRChatDLSS5Cam-win64.zip from the latest release and extract it to any location. (Radeon card: VRChatDLSS5Cam-win64-amd.zip instead, see AMD Radeon cards below.)

2. Nothing else has to be copied. The DLSS 5 runtime is in the archive under runtimes\, in one build for RTX 50 and one for RTX 40 / 30 / 20; the application picks the build for the installed card at start and switches to the other one when that fails. The build in use is named next to Runtime in the DLSS 5 section, and the path is recorded in log.txt.

   VRChatDLSS5Cam.exe
   runtimes\blackwell\nvngx_dlssnr.dll   RTX 50
   runtimes\universal\nvngx_dlssnr.dll   RTX 40 / 30 / 20
   runtimes\other\nvngx_dlssnr.dll       a runtime for another vendor (not included; Radeon cards use the Radeon edition)
   

   A file of your own can be used instead: nvngx_dlssnr.dll next to VRChatDLSS5Cam.exe is tried after the build for the card, and any file can be selected under DLSS 5 Neural Rendering → Runtime path.

3. In VRChat, open the Camera, switch it to Stream mode and enable Spout Stream in its settings.

4. Start VRChatDLSS5Cam.exe. The first start opens a short setup guide (language, GitHub access, what the controls do, where to find help). The camera picture then appears in the preview with DLSS 5 applied, and the badge next to the Enable DLSS 5 switch reads Active.

5. Frame the shot in VRChat and press Ctrl+Alt+P (or use the Capture photo button). The PNG is written to Pictures\VRChat DLSS5 Cam.

While nothing is open, the preview shows the Get started page: the three steps (choose the source, adjust DLSS 5, save) and a tile for each way in. Live from VRChat switches to the live camera and says whether a Spout sender is there; Open a picture and Open a video open a file. A file dropped anywhere on the window opens as well, and several files or a folder go to the library.

Notes on the live camera

• The stream resolution is decided by VRChat. Raising camera_spout_res_width / camera_spout_res_height in VRChat's config.json produces a sharper input, and the application adapts to it automatically.
• When the neural pass is too slow for a live preview on a given card, Neural pass only for captures in the DLSS 5 section keeps the preview on the plain picture with the neural, motion and depth passes at rest, and runs the neural pass for each capture only.
• Processing rate cap in the Display section limits how many camera frames per second are processed, which keeps the GPU free for VRChat.
• Both are expert controls: they are shown while the Advanced switch at the top of the sidebar is on (the default).

AMD Radeon cards

The Radeon edition, VRChatDLSS5Cam-win64-amd.zip, runs the neural pass on Radeon cards through DLSS-NR-on-AMD, a separate project by Daniel Blanco that brings DLSS 5 neural rendering to AMD hardware by attaching to programs that use FSR. Nothing of that project is included here: the edition hosts an FSR context (AMD's FidelityFX API, amd_fidelityfx_dx12.dll, MIT) for it to attach to, and fetches its installer from that project's own release page when you press the button. That project's licence allows personal, non-commercial use and no redistribution or modification, which is why its installer is never part of this archive.

1. Download VRChatDLSS5Cam-win64-amd.zip from the latest release and extract it to any location. It carries amd_fidelityfx_dx12.dll and the RTX 50 build of nvngx_dlssnr.dll next to the executable (the installer of step 2 converts that file into its own weights); the NVIDIA-only files (nvngx_dlss.dll, runtimes\) are not in it. The GeForce edition offers Get the Radeon edition… when it finds itself on a Radeon card (--edition amd on the command line does the same) and swaps its files the way an update does.
2. Start VRChatDLSS5Cam.exe. The Get started page and the DLSS 5 section offer Install DLSS-NR-on-AMD…, with the licence linked next to it. One press downloads dlssnr_on_amd_setup.exe into the program folder and runs it in the background, taking the folder and the DLL name it proposes for the executable; no window and no key press are needed. When it has finished, the application restarts by itself, and the DLSS 5 section reads Runtime: Loaded FSR … · FSR host and DLSS-NR-on-AMD v…: loaded (version.dll) · supported release, with the badge next to Enable DLSS 5 reading Active. If the folder needs administrator rights, or the card is one the port does not support, the installer opens its own window instead so its message is visible.

The DLSS 5 section shows the installed release next to the one this version works with, DLSS-NR-on-AMD v0.2.17. Its later releases ship as NeuralHub, a different installer with a different layout on disk, which this version does not drive yet; the Install button fetches v0.2.17, and an installation made with it keeps working. Support for NeuralHub will come in a later version. When the installation is older than v0.2.17, the section says so and the button reads Update to …; it and Run the installer again… fetch the installer and open its window, where U updates (keeping the settings and weights) and R removes it. An installation made before this version is recognised by the installer file next to the executable.

On this route the strengths of the neural pass up to 1 belong to DLSS-NR-on-AMD and are set in its own overlay (End key; its Mode there stays at inline, which a saved frame needs, see below); this application's Style and strength values are not passed to it. The controls that act after the pass (Output blend, strengths above 1, Neural pass resolution) work as usual. DLSS-NR-on-AMD needs Windows 11, a Radeon RX 7000 or RX 9000 card and Adrenalin 26.1.1 or newer; its release page states the current requirements. The Radeon edition always runs the FSR host; the GeForce edition hosts nvngx_dlssnr.dll itself. There is nothing to choose, and each edition points to the other one when it finds the other vendor's card.

The Radeon edition has been run on an RX 9060 XT: still pictures at 720p and 4K, video files, batch processing from the command line and the live VRChat camera (at a lower frame rate than on a GeForce card; see the list below). DLSS-NR-on-AMD's installer sets an inline budget of 200 ms, suited to a game, after which a frame shows the previous frame's result; the application sets InlineWaitMs in dlssnr_on_amd.ini to 500 after the install (or at start-up, with one automatic restart, when the file has another value) so that a saved picture or video frame carries its own result, and keeps Inline at 1 there. Nothing else in that file is changed. log.txt (%LOCALAPPDATA%\VRChatDLSS5Cam\) lists the DLSS-NR-on-AMD files next to the executable and ends with the last lines of that project's own log (dlssnr_on_amd.log in the program folder); reports with it are welcome in the issues.

What to expect on Radeon in this version

• Cards. Radeon RX 7000 and RX 9000 on Windows 11 with Adrenalin 26.1.1 or newer, the cards DLSS-NR-on-AMD supports; the author's tests ran on an RX 9060 XT. On an older Radeon the FSR host still runs, but without the port's network, and the picture comes back unchanged.
• Works. Still pictures, video files, the batch queue, the live VRChat camera with captures and timelapses, the depth network, the controls after the pass (Output blend, strengths above 1, Neural pass resolution), updates and the edition switch.
• Slower. DLSS-NR-on-AMD runs its network inline and the frame waits for it: about 25 ms at 720p and 180 ms at 4K on the RX 9060 XT, so a 4K live picture reaches a few frames per second. A Maximum resolution cap under Neural pass resolution and the Processing rate cap keep the live picture fluid; saved pictures and videos come out the same, only later.
• FSR optical flow in place of the hardware engine. The optical flow engine belongs to the GeForce driver and has no counterpart on Radeon. Since 1.8.0 the motion vectors come from the FSR optical flow instead, a pyramid of compute passes of this program's own in the shape of the optical flow of the AMD FidelityFX SDK (MIT): it follows fast motion far better than plain block matching and checks its vectors in both directions. It is what the Radeon edition starts with, and Search radius and the bidirectional check under Frame guidance trade its GPU time against quality.
• Strengths and presets. The port keeps its own Preset, Style and strengths up to 1, set in its overlay (End key), as described above. Moving those sliders here therefore composites the existing result at once (values above 1 take effect in the composite) instead of running the picture through its passes again, which on a Radeon costs about four seconds at 4K for no change in the picture.
• The depth network and the port take turns. The port's network and the depth network (DirectML) do not share the card: a neural pass that ran alongside the estimator's warm-up stretched from 0.17 s to 2.7 s, past the driver's two-second limit, and the graphics device was lost (the driver restarts, the program has to be started again). Since v1.5.0 the program waits for the estimator before every neural pass, so the two never overlap; the first passes of a picture start once the estimator is ready.
• A stalled network job. A DLSS-NR-on-AMD job that does not end in time holds the frame until the port's wait gives up, 2 to 2.5 times InlineWaitMs later, and that frame shows the previous frame's result. With the 1000 ms earlier versions set, the wait reached the driver's two-second limit and a video could end with the graphics device lost; with 500 ms it stays near 1.3 s, and the next frame waits until the stalled job has ended.
• Idle load. Once a picture has run its passes the card idles (about 2 % on the RX 9060 XT, the same as on a GeForce). One CPU core, though, stays busy from the first neural pass until the program closes: a thread of DLSS-NR-on-AMD polls the card at full speed (its inline mode). That thread belongs to the port; releasing the FSR context does not end it.

Images and video files

A file can be dropped onto the window or opened with Open image… / Open video… in the Source section. The picture appears in the preview and every slider takes effect on it immediately. The ✕ button beside Open… closes the file again: processing stops, the preview empties and the file stays in the library. Process & save PNG (images), Process & save video (videos) or the hotkey writes the result into the save folder of the Capture section (Pictures\VRChat DLSS5 Cam unless changed; the source file's own folder with Save next to the source file on) as <name>_DLSS5_<w>x<h>.png, .mp4, .gif / .png (APNG) / .webp for an animation, or a folder of PNG frames.

For video, the controls under the preview play and pause the file through the whole pipeline, step one frame at a time, and show a thumbnail of the frame under the cursor on the seek bar. Start here / End here limit the processing (and the audio) to a range, and Whole video clears it. By default the output matches the source: same codec, frame rate and bitrate. With Match the source switched off, H.264, HEVC, a PNG sequence, GIF, APNG or WebP and the bitrate can be chosen manually. The Capture section shows roughly how long the file will take with the current settings.

An animated GIF, APNG or animated WebP opens as a video: the same play, step and range controls apply, and every frame keeps its own display time. With Match the source the result is written in the same format with the file's loop count (a lossless WebP stays lossless); with it off, any moving source can be saved as MP4, a PNG sequence, GIF, APNG or WebP, and WebP quality sets the compression (100 = lossless). Transparency is kept when Keep transparency (alpha) is on (GIF: one bit). A GIF holds 256 colours per frame and at most 50 frames per second.

Batch processing

Every file that is opened or dropped is collected in the library below the preview; Add files… and Add folder… add more; while it is empty, the library is a single row with those buttons and opens with the first file. A double-click on a thumbnail opens that file in the preview for adjustment. A single click selects it alone, Ctrl+click adds or removes one file, Shift+click extends the selection from the file clicked last, dragging on empty space draws a selection rectangle, and a click on empty space clears the selection. The box that appears at a thumbnail's corner under the mouse (and on every thumbnail while any is selected) ticks it, the ✕ at its other corner takes it out of the library, and Ctrl+A or Select all in the … menu selects every readable file. The Delete key or the Delete button takes the selected files out of the library while the mouse is over it (or after the last click was inside it); the files themselves stay on disk. Process selected or Process all then runs the files one after another, with a bar on each thumbnail showing the progress and the state remaining visible afterwards. In a narrow window the header's buttons shrink to icons, and those that no longer fit move into the … menu.

The right mouse button opens a menu on a thumbnail: show the file in Explorer, take it out of the library, or give it its own DLSS 5 parameters in a separate window. With several files selected the same menu offers Own parameters for N files…, which sets the values for all of them at once; each file keeps its own copy afterwards.

A small toolbar over the preview of a library file turns, mirrors and crops it: a quarter turn to the left or right, a horizontal or vertical mirror, or a crop frame whose corners and edges are dragged (dragging inside the frame moves it), confirmed with Enter or Apply and dropped with Esc. The last button restores the file to its original state. Each file keeps its own orientation and crop, processing and the saved result use the turned and cropped picture, and every step is recorded in the history and can be undone. The live picture has the same toolbar without the crop: a turn or mirror of the VRChat camera picture applies to the preview, the captures and the timelapse alike, and is kept across sessions. The toolbar itself can be moved: drag it by the grip at its left end, and drop it past an edge of the picture, or press the arrow at its right end, to tuck it away at that edge, where a small tab brings it back. Its place is kept across sessions as well.

Working with the interface

• Three sections cover the everyday work: Source (what comes in), DLSS 5 Neural Rendering (how it looks) and Capture (where it goes), followed by Display, MCP and About; a click on a section's title folds it. The Advanced switch at the top of the sidebar, beside the search field, shows or hides everything beyond the essentials: the tone and structure sliders, the neural pass resolution and the output blend, the neural pass readouts, Neural pass only for captures, Hardware decoding, Keep transparency (alpha), the Frame guidance and DLAA pre-pass sections, V-Sync and the processing rate cap, the MCP section's queue limits, Internals with the timers, and the pass timings in the status bar. It is on by default; switched off, the sidebar keeps only what a first picture needs.
• The top bar holds the source switch (Live, Picture, Video) with a badge for its state and a DLSS 5 badge while the neural pass runs; on the right undo, redo, the history, help (Documentation, a book with a question mark, and Ask AI, a speech bubble with AI, to its right), the language and the main action: Capture photo, Process & save PNG or Process & save video. The status bar at the bottom names what is open with its size and frame rate, the progress of a running job with the time left, and the file saved last, with a button that shows it in Explorer.
• Every label fits. Long names wrap onto a second line instead of being cut, a value or badge with no room left moves to the next line, and a button pair that no longer fits side by side is stacked, in all four languages and down to the narrowest sidebar.
• Comparison is done with the wipe (the handle in the preview is dragged), the original view, or the depth and motion-vector views. The wipe follows the mouse at the interface's own frame rate, even while a video plays slowly. The mouse wheel over the preview zooms around the cursor, dragging pans, and a double-click returns to the fitted view.
• Presets are created in the DLSS 5 section: + saves the current values under a name, the list applies one with a click, and every entry can be overwritten with the current values, renamed or deleted. They are kept in presets.txt in the settings folder.
• Search over the settings is available from the field at the top of the sidebar: only the matching controls and their sections stay visible. Esc or the ✕ clears it.
• Undo and redo cover every change to the settings and every file added to or removed from the library, through Ctrl+Z / Ctrl+Y or the two arrows in the top bar. The clock button beside them opens the History: every recorded change as a list (Intensity: 1.2, Added photo.png, Removed 3 files, photo.png: own values), with the current state highlighted and the undone states listed dimmed below it. A click on an entry jumps to that state, and the later ones remain available until the next change is made, so the history can be traversed in both directions. Up to 100 steps are kept.
• Fullscreen is entered from the button in the corner of the preview or with F11: the picture alone, over the whole screen. The video controls appear while the mouse moves, and Esc or F11 leaves. Entering and leaving fade, and so does the switch between Live, Picture and Video.
• Dark or light. Theme in the Display section: System follows the Windows app colour setting, or Dark and Light can be selected directly.
• The sidebar slides away when the gap between it and the preview is clicked (a small arrow in the middle of the gap shows the direction), and the gap above the library folds the library the same way. The thin line along the gap's edge — the one that turns blue under the mouse, where the pointer becomes a resize arrow — is the drag target for the sidebar's width or the library's height, and the thumbnails grow or shrink with it. Both sizes are remembered for the next start. While a file is being processed the sidebar is locked and offers Cancel.
• Help is one click away: the Documentation button in the top bar (or Documentation in About) opens the documentation site in the interface's language, and Ask AI beside it opens a panel at the right edge that answers questions from that documentation (see AI Q&A).
• At start a small card with the icon, the name, the version (marked Pre-release on pre-release builds), a status line and a moving bar shows what the application is doing; the window itself fades in only once its first frame is ready, so no blank window appears at the start. The file from the last session is opened again only if Reopen the last file at start in the Display section is on (it is off by default).

Keyboard and mouse

Keys	What happens
Ctrl+Alt+P	Capture a photo of the live camera, or process the open image or video (global hotkey, changeable in the Capture section)
Ctrl+Z · Ctrl+Y / Ctrl+Shift+Z	Undo · redo a settings change or a library change (the clock button in the top bar lists them all)
F11 · Esc	Fullscreen preview · leave it
Esc in the Ask AI panel	Close a notice in the panel if one is showing, else the panel (the conversation stays until the application closes)
Ctrl+A · Delete	Select every readable file in the library · take the selected ones out of it, keeping the files on disk (mouse over the library)
Space	Play / pause the open video
← → (Shift: 10 frames) · Home End	Step through the video · jump to the ends
I · O	Set the start · the end of the range to process
Mouse wheel over the preview · drag · double-click	Zoom · pan · back to the fitted view
Click a thumbnail · double-click · Ctrl+click · Shift+click · right button	Select only that file · open it in the preview · add or remove one · extend the selection · open the menu
Drag on the library's empty space · click on it	Selection rectangle · clear the selection
Click the gap beside the sidebar or above the library · drag the thin line at its edge	Fold or unfold that panel · change its width or height
Drag the left edge of the Ask AI panel	Change the panel's width (remembered)
Enter · Esc while cropping	Apply the crop · cancel it

Updates

At every start the application asks GitHub for the newest release on the selected channel and reports it when it is newer than the running version. Two channels are offered in the About section under Update channel: Stable (full releases only) and Pre-release (also the builds published for testing before a full release; full releases reach this channel too, and the update window marks each release Full release or Pre-release). Check for updates at start switches the check off, the Check for updates button runs it at any time, and the result of the last check is shown underneath.

Where GitHub is slow or unreachable (mainland China, for one), GitHub access in the same section switches the check and the download to a mirror site: Mirror sites, the fastest one measures eight known sites and uses the one that answers fastest (the Measure the sites button shows the result of each), A mirror site of my own takes a site of your own, and GitHub directly talks to GitHub itself. When no site answers, a window offers the other choices. Through a mirror the program reads the release list updates.json that this repository keeps in step with the releases. In the Radeon edition the look-up of the DLSS-NR-on-AMD release and the download of its installer go through the chosen site as well; the entry comes from port.json in this repository, which names the release the program works with, so a change there reaches every installed copy without a new version. The program reads notice.json from the repository the same way: the project's notices, shown as a card at the top of the sidebar until dismissed, for word that cannot wait for a release (such as a change at DLSS-NR-on-AMD).

When a newer version exists, a window shows its version, its date and its release notes with three buttons. The notes come in the interface's language (English, Chinese, Japanese or Korean; in English when a release has no translation). Update now downloads the release zip into %LOCALAPPDATA%\VRChatDLSS5Cam\update, unpacks it, closes the application, replaces the program files and starts it again; the settings and the library are left alone, the bundled runtime builds under runtimes\ are replaced with the release's, and a runtime file of your own next to the executable stays in place. Release page opens the release in the browser and Later closes the window. The program folder has to be writable, which a folder under Program Files usually is not; the application says so, and the update can be applied by hand from the release page instead. A failed check or a failed update is shown as a notification.

Apart from two features that act only when you use them, the application's network access is limited to GitHub: api.github.com for the release list, github.com for the downloads and raw.githubusercontent.com for updates.json, port.json and the translated release notes, or the chosen mirror site in their place; the Radeon edition also asks DLSS-NR-on-AMD's release page. The two features: AI Q&A talks to Mintlify (and hCaptcha) once Ask AI has been opened, and sends what is typed into it, never your files (see AI Q&A); the MCP server answers only the clients you allow, and downloads a picture or video only when a client's job names it by a web address (see MCP). Nothing else is sent anywhere. Headless and --process runs never check on their own.

AI Q&A

Ask AI in the top bar, beside Documentation, opens AI Q&A in a panel at the right edge of the window; the picture and the sidebar move over to make room. It is the AI Q&A of the documentation site, provided by Mintlify: a question typed there gets an answer written from the documentation, with links to the pages it used. The panel follows the interface's language and theme (a change of language starts a new conversation), links open in the default browser, and links to the documentation open in the interface's language. Esc in the panel, Ask AI again or the panel's ✕ closes it; closing only hides it, so the conversation stays until the application closes. Its left edge is a drag line like the sidebar's: the panel is at least 16 times the font size wide and always leaves the picture 20 times the font size, and its width is remembered (askWidth in settings.ini, in font sizes). The answers are written by AI and can be wrong, as the line at the top of the panel says: the page an answer links to is the reference. Mintlify's own controls in the panel (Clear chat, Found results for …) stay in English.

• From an answer to the control. Names of the interface in an answer (in bold, as code or in quotes, as the interface writes them) get a dotted underline. A click shows the place in the application: the sidebar comes out, the section opens, the sidebar glides to the control and a ring lights up around it; nothing is set or changed. Pointing at a name offers Show in the app or Open its page in the documentation. A control that Advanced hides is shown through the settings search, one that appears only in some states points at the control that decides it, one of another source rings that source's button in the top bar, and when nothing can be shown the control's page of the documentation opens.
• Online, with an allowance. AI Q&A runs on Mintlify's servers and needs an internet connection; when the panel cannot load, it says AI Q&A could not be loaded and offers Try again and Open documentation. The project has a monthly allowance of answers, shared by everyone who uses the application; once it is used up, AI Q&A stops answering until the next month. It is most useful for finding the right page; for a single control, the documentation's settings page or its search is quicker, and the documentation is always there.
• When a question gets no answer, a notice above the input box says why, in the interface's language, with the way on: AI Q&A is unavailable right now (the allowance is used up or the service is down; Try again / Open documentation), Too many questions in a short time and Your question could not be sent (Open documentation / Try again), This conversation has reached its length limit (Open documentation / New conversation). Esc closes the notice first and the panel with the next press.
• Nothing before the first opening. Nothing of it starts, not even the WebView2 control that draws the panel, until Ask AI is clicked for the first time, and nothing is sent before then.
• What is sent. What is typed into the panel goes to Mintlify, which writes the answer; Mintlify also receives the panel's usage events and checks questions with hCaptcha. Pictures, videos, settings and the library never leave the computer; a file attached in the panel by hand would go to Mintlify too.
• What is kept. The panel's cookies and storage are kept in %LOCALAPPDATA%\VRChatDLSS5Cam\webview2 (or in webview2 inside the --data-dir folder); with the application closed, the folder can be deleted. log.txt records the panel's opening and closing, its errors, and the HTTP status and length of each answer, never the questions or the answers' text.
• Without WebView2. The panel needs the Microsoft Edge WebView2 Runtime. Where it is missing or cannot start, and under Proton (the Linux package), Ask AI opens the documentation site's AI Q&A in the browser instead (?ask in the address) and the application says AI Q&A opened in your browser.

MCP

An MCP client (Claude Desktop, Claude Code, Cursor, ...) can work the program the way you do: open files, change every setting, run the library, save captures and look at the preview. Turn on Run the MCP server in the sidebar's MCP section, or let the client start the program itself with this block in its MCP configuration (the path is the installed executable's; the sidebar's Copy client configuration button writes it for you):

{ "mcpServers": { "vrchat-dlss5-cam": { "command": "C:\\Path\\To\\VRChatDLSS5Cam.exe", "args": ["--mcp"] } } }

Claude Code: claude mcp add --transport http vrchat-dlss5-cam http://127.0.0.1:51550/mcp while the program runs with the switch on. By default the server listens on 127.0.0.1 only and connects nowhere by itself (the one download it makes is a job's input given as a url); Read only limits clients to looking.

Other computers can send work too: set Reach to Local network, add a key per bot or person (roles Jobs, Admin, Viewer), and a chat bot or a script on another PC uploads a picture or a video, is told at once its place in the queue and when its turn comes, and downloads the result later. One PC with a strong card can serve several bots this way; the sidebar shows the running job and can cancel it. Every tool, the queue's rules and a minimal bot are described in docs/MCP.md.

Settings reference

All settings, section by section

Section	Setting	Meaning
Source	Input	VRChat camera (Spout), Image file or Video file.
Source	Spout sender	Which Spout sender to receive; VRChat's camera is VRCSender1.
Source	Custom output resolution	Output at a chosen size instead of the source size. Larger than the source = upscaling: DLSS super resolution (the official DLSS upscaler rebuilds the detail from the source at a render size chosen for the ratio, then the neural pass works on the large picture) or plain Resampling. Super resolution multiplies the pixels every later stage has to process: expect a large drop in speed and more video memory in use; for a live stream set a processing rate cap. Smaller = downscaling. DLSS works up to 8192 pixels a side and the neural pass up to about 45 megapixels: a larger output is finished within those limits and resampled to the requested size.
Source	Match the source	Videos only, on by default: the output uses the codec (H.264 or HEVC), the frame rate and the average bitrate of the source file. An animated GIF, APNG or WebP comes back in the same format with its loop count.
Source	Save as / Bitrate / Keep audio / WebP quality	With Match the source off: MP4 (H.264), MP4 (HEVC), PNG sequence, GIF, APNG or WebP; the encoder bitrate (5–200 Mbit/s) and whether the audio track is copied for MP4; the compression of an animated WebP (50–100, 100 = lossless).
Source	Hardware decoding	Decode the video on the GPU. Best switched off when a file shows wrong colours or fails to open.
Source	Paper white / Highlight compression	Shown only for floating-point (HDR) Spout textures: exposure reference and soft highlight roll-off before the neural pass.
DLSS 5	Enable DLSS 5 (DLSSNR)	Switches the neural pass on or off. Off releases the runtime; on loads it again from the file.
DLSS 5	Runtime path / Reload	The runtime file in use; empty means the bundled build for the installed card. Reload starts that choice over and loads the file again.
DLSS 5	Preset	Named sets of the DLSS 5 values: + saves the current ones, the list applies one, each entry can be overwritten, renamed or deleted. Kept in presets.txt in the settings folder.
DLSS 5	Style	Render style (default / natural / cinematic) passed to the runtime.
DLSS 5	Intensity	Overall strength of the neural pass, 0–2. Up to 1 it is the runtime's own strength; above 1 the application amplifies the difference between the neural result and the original (which can exaggerate artifacts). At 0 the picture is left untouched.
DLSS 5	Global tone strength / Local tone strength	0–2, with the same rule above 1 (the highest strength above 1 sets the gain).
DLSS 5	Local structure strength / Skin structure strength	Detail enhancement, 0–2, same rule above 1. Skin structure may be left at the runtime default.
DLSS 5	Auto mask / UI correction	Automatic subject mask, UI-safe processing.
DLSS 5	Neural pass only for captures	For cards too slow for live use: between captures the neural, motion and depth passes all rest and the GPU stays nearly idle; a capture first runs the neural pass for 16 fresh frames (after the first depth estimate when the depth network is in use: DLAA, super resolution or the depth view), then saves. Still images are not affected.
DLSS 5	Input exposure (paper-white scale) / Tone transfer strength / Colour strength	Output blend. Input exposure (0.25–4×) scales the picture the network sees and is undone afterwards. Tone transfer and colour strength (0–2) set how much of the neural pass's brightness and colour changes reach the output; 1 / 1 reproduces the neural result exactly, 0 keeps the original.
DLSS 5	Shadow strength / Highlight & glow strength	Output blend, 0–2: how much of the neural pass's darkening and of its brightening reaches the output. 1 / 1 = as rendered.
DLSS 5	Neural pass resolution	Sizes the neural pass: the full picture, a cap on its long edge (a large source, for example 8K, is processed at a fixed, smaller size) or a percentage of the picture. A reduced pass has its change upsampled onto the full-resolution picture; lower values cut the GPU load at the cost of the finest detail.
Capture	Folder / Save next to the source file / File name (Output file name with a file open) / Keep transparency (alpha) / Also save the original frame / Global hotkey / Auto capture every (seconds)	Where and how photos and videos are saved. With Save next to the source file on, a processed picture or video goes into its source's folder instead of the capture folder. The file name comes from a template: {name} (the source file's name, VRChat for a live capture), {date}, {time}, {size}, {width}, {height}, {insize}, {inwidth}, {inheight}; anything else is kept as typed, and a name already taken gets _2, _3, … Live captures default to VRChat_DLSS5_{date}_{time}_{size}, processed pictures and videos to {name}_DLSS5_{size}.
Capture	Estimated time	Rough processing time of the open image or video with the current settings, refined by every run.
Frame guidance	Motion vectors	NVIDIA Optical Flow (with a forward/backward consistency check), the FSR optical flow (this program's own pyramid of compute passes, on any card, with its own search radius and consistency check), GPU block matching, or none.
Frame guidance	Depth	AI estimated (Depth Anything V2 Small on DirectML; update interval and network resolution adjustable), flat, gradient, or zero. DLSS 5 itself does not read the depth; the network is loaded only while DLAA, super resolution or the depth view needs it.
Frame guidance	Auto reset on scene cut	Clears the temporal history on scene cuts. Off by default.
DLAA pre-pass	Run DLSS DLAA before neural rendering / DLSS preset	Optional DLSS anti-aliasing pass at native resolution before neural rendering. While DLSS super resolution is in effect it already includes this pass; the preset applies to both.
Display	Theme	System (follows Windows), Dark or Light.
Display	Compare / Fit to window or 1:1 / Zoom / Checkerboard behind transparency / Show the library / Show overlay info / Show log / V-Sync	Preview options.
Display	Reopen the last file at start	Open the file from the previous session again at the next start. Off by default.
Display	Processing rate cap	Live camera: process at most this many frames per second and skip the rest. 0 = every frame.
About	Check for updates at start / Update channel / Check for updates	Look for a newer release at every start (on by default), on the Stable or the Pre-release channel, or right away with the button. The result of the last check is shown underneath.
About	Check the graphics driver at start	On a GeForce card, compare the driver with the oldest version the bundled DLSS 5 runtime runs on (616.56) at every start and show a notice when it is older (on by default; Don't show again in the notice turns it off). The DLSS 5 section keeps its own warning line either way.
About	GitHub access	GitHub directly, through the fastest of eight known mirror sites (Measure the sites shows how each answered), or through a mirror site of your own. For regions where GitHub is slow or unreachable.
About	Open log file / Open settings folder / Report a problem / Documentation / Project page / Setup guide / Third-party notices / Reset all settings	Version, GPU and driver, the issue form on GitHub (prefilled with the version and the graphics card), this guide, the setup guide (the pages the first start shows: language, GitHub access, what the program does, what the settings do, where to find things), and the maintenance buttons.

Settings are stored in %LOCALAPPDATA%\VRChatDLSS5Cam\settings.ini; the log is log.txt in the same folder, and the logs of the five sessions before it are kept there as log-1.txt to log-5.txt (log-crash.txt after a crash).

Command-line options (open files, process unattended, screenshots, headless runs) are listed in docs/COMMAND_LINE.md.

Troubleshooting

• "No Spout sender found" / "Waiting for VRChat Spout stream…" – Spout Stream has to be enabled on VRChat's Stream camera, and the camera must be open. The first appears while no Spout sender runs at all, the second while senders run but the chosen one has not sent a picture yet. Other Spout senders are listed in the Spout sender box.
• "The runtime files are missing" – the archive was not extracted completely: runtimes\blackwell\nvngx_dlssnr.dll and runtimes\universal\nvngx_dlssnr.dll belong next to VRChatDLSS5Cam.exe. Extract it again, or select a runtime file under Runtime path.
• Graphics driver too old – the bundled DLSS 5 runtime needs GeForce driver 616.56 or newer; on an older one the neural pass does not start (log.txt: PlatformError). The application checks the driver at every start and, when it is older, says so in a notice with NVIDIA's download page; Don't show again turns the notice off and Check the graphics driver at start in About turns it back on. The DLSS 5 section keeps a warning line until the driver is updated; the viewer, captures and videos work on the old driver.
• Neural rendering failed – the application switches to the other bundled build by itself and says so in a notice. When neither build starts, the message under the error says so; a newer graphics driver is the first thing to try, and log.txt carries the NGX result code. On a Radeon card the neural pass runs through the Radeon edition and DLSS-NR-on-AMD instead (see AMD Radeon cards).
• FSR host runs, but the picture is unchanged – DLSS-NR-on-AMD is not attached to the process: the DLSS 5 section reads DLSS-NR-on-AMD: not loaded. Install it there (Install DLSS-NR-on-AMD…); the application restarts by itself once the installer is through. When it is installed but not loaded, run its installer again. When it is loaded and nothing changes, open its overlay (End key) to see that it is enabled and try the other Mode; its own log (dlssnr_on_amd.log in the program folder; its last lines also end log.txt) says what it did. The port is a separate project with its own requirements.
• amd_fidelityfx_dx12.dll is missing – the FSR host route needs that file next to the executable; the Radeon edition (VRChatDLSS5Cam-win64-amd.zip) carries it.
• Neural pass active, but the picture is black or unchanged – every neural frame is compared with its input, and a warning appears when the runtime reports success but delivers a black or unchanged picture (log.txt: "DLSSNR output check"). Switching DLSS 5 off and on again starts it fresh. If it persists, that runtime build does not produce a picture on this GPU. (With Intensity at 0 an unchanged picture is normal and no warning is shown.)
• The application does not start / closes immediately / crashed – %LOCALAPPDATA%\VRChatDLSS5Cam\ holds log.txt (its last line is the step that failed), the logs of the five sessions before it (log-1.txt to log-5.txt) and crash.txt. At the next start the application notices that the previous session did not end properly, keeps that session's log as log-crash.txt and offers the log folder and a prefilled issue form (Report on GitHub; the About section has Report a problem at any time). log-crash.txt and crash.txt belong in the issue report.
• NGX not initialized / DLAA unsupported – the NGX runtime needs an NVIDIA GPU and a current driver. It serves DLAA and DLSS super resolution only; the DLSS 5 neural pass does not depend on it and keeps working.
• Depth estimator unavailable – onnxruntime.dll, onnxruntime_providers_shared.dll, DirectML.dll and models\depth_anything_v2_small_fp16.onnx must sit next to the executable (all are in the release package). Until the estimator is ready the application uses zero depth; its state is shown under Frame guidance.
• Optical flow unavailable – on a Radeon card this is expected: the optical flow engine is part of the GeForce driver and has no counterpart on Radeon, so the FSR optical flow is used and the status dot under Frame guidance says so. On a GeForce card, when log.txt says "NVOF unavailable, falling back to the FSR optical flow", the GeForce driver needs an update; the FSR optical flow is used until then.
• The video preview is black – many films start with a fade from black; while the frame on show is almost black, the preview says so over the picture. Seeking forward with the bar or the arrow keys moves past those frames.
• Video file does not open / no encoder available – the formats depend on the codecs installed in Windows. HEVC files need the HEVC Video Extensions (Microsoft Store), and Windows N / KN needs the Media Feature Pack. When the H.264 encoder is missing, PNG sequence is the alternative output. Switching Hardware decoding off helps with files the GPU decoder rejects.
• Low frame rate – switch DLAA off, raise the depth update interval or lower the depth network resolution, lower the neural pass resolution, or set a processing rate cap. The optical-flow grid is best left at 4 px (2 px and 1 px cost far more at 4K). The log prints a Perf: line every 15 s with the cost of each stage.
• The graphics device was lost – the graphics driver reset the card. Windows gives a GPU job about two seconds, and a card that is fully loaded for a long time can miss that: typically a live stream with VRChat rendering on the same card while the neural pass upscales every frame (log.txt: "Device removed", DXGI_ERROR_DEVICE_HUNG). Every program on the card loses its picture at that moment. Since v1.5.1 the application starts again by itself once the driver is back, says so in a notice, and keeps the log of the lost session as log-device-loss.txt next to log.txt; its last Perf: line shows the load and the video memory in use. If it repeats, lower the load: a processing rate cap, a lower Neural pass resolution or a maximum-resolution cap, a smaller output resolution (super resolution off), or a frame rate limit in VRChat; a newer graphics driver is worth a try. Scripted runs (--process, --headless, --exit-after) exit with code 2 instead.
• A video exported without sound – a file with several audio tracks (an iPhone MOV carries a stereo AAC track and a spatial audio track that Windows cannot decode, listed first) used to take the first track only. Since v1.5.1 the first track that decodes is used; log.txt names the tracks and the one taken.
• The update cannot be installed – the program folder has to be writable, which a folder under Program Files usually is not. Moving the application to a user-owned folder, or downloading the new version from the release page and replacing the files by hand, both work; settings and library stay where they are.

Sharing results

Photos and videos finished with the application are welcome on X under the hashtag #VRCDLSS5CAM, which also collects the shots posted by others. Bug reports and feature requests are handled through GitHub Issues.

Building from source

Requirements: Visual Studio 2022 (MSVC v143, Windows 10 SDK) and CMake 3.21+.

git clone https://github.com/AlanBacker/VRChat-DLSS5-Cam.git
cd VRChat-DLSS5-Cam
cmake -S . -B build -A x64
cmake --build build --config Release --parallel

The configure step downloads the NVIDIA DLSS SDK (headers, nvsdk_ngx_s.lib, nvngx_dlss.dll) from NVIDIA's public GitHub repository, ONNX Runtime (DirectML build), DirectML and the Microsoft Edge WebView2 SDK (its static loader, built into the executable for the AI Q&A panel) from NuGet, libwebp 1.6.0 from GitHub (built into the executable), and the Depth Anything V2 Small FP16 model from Hugging Face (-DVDC_FETCH_DEPTH_MODEL=OFF skips the model). All downloads are hash-checked. Shaders are compiled at run time, so no shader toolchain is needed. The FidelityFX SDK v1.1.4 headers and its signed amd_fidelityfx_dx12.dll (MIT) are downloaded from AMD's GitHub repository in the same way. -DAPP_EDITION_AMD=ON builds the Radeon edition (FSR host route, no NVIDIA-only files in the package); the default is the GeForce edition. The DLSS 5 runtime is not part of the build: the release workflow places its two builds into the archives from the runtime-310.8 release.

Architecture

Pipeline and design notes

VRChat Stream Camera ──Spout──▶ D3D11on12 receive ──▶ convert (sRGB / resize)
      ▶ NVIDIA Optical Flow (forward + backward) / FSR optical flow (pyramid, both directions) / block matching ──▶ motion vectors + confidence
      ▶ Depth Anything V2 (ONNX Runtime DirectML, every N frames) ──▶ normalized depth, reprojected in between (for DLSS SR / DLAA and the depth view)
      ▶ [DLSS SR / DLAA] ──▶ DLSSNR (nvngx_dlssnr.dll) ──▶ composite / compare ──▶ preview + PNG capture
Video file ──Media Foundation──▶ decode (GPU) ──▶ same pipeline, one frame at a time ──▶ MP4 (H.264 / HEVC + AAC) or PNG sequence
Animated GIF / APNG / WebP ──WIC / libwebp──▶ decode ──▶ same pipeline, one frame at a time ──▶ GIF / APNG / WebP (frame timing kept), MP4 or PNG sequence

The application hosts the DLSS 5 neural-rendering snippet outside the NGX runtime: the DLL is loaded directly, its module-name check is satisfied, and the DLSSNR.* NGX parameter contract is used to create and evaluate the feature on a D3D12 queue (src/ngx/DlssnrFeature.cpp). The parameters are the same set the RenoDX DLSS 5 add-on exposes.

The guidance scheme is built for video input: same-resolution SDR input, hardware optical flow whose confidence is lowered where forward and backward vectors disagree, monocular depth from Depth Anything V2 normalized (2nd/98th percentile) to inverted relative depth and carried along the motion vectors between inferences, and no per-frame history resets. Everything is an independent MIT implementation (src/gfx/Pipeline.cpp, src/gfx/DepthEstimator.cpp, src/gfx/Shaders.cpp). The optical flow engine is driven through its D3D12 interface on the program's own device; each pass is ordered against the graphics queue with fence points, so a frame always reads the vectors computed for it (src/gfx/NvOpticalFlow.cpp).

The FSR optical flow is this program's own and needs no engine: the coarsest level of a luma pyramid searches for the large motion, every level below tries the vectors it inherits together with those of its neighbours and refines around the best, a median over whole vectors runs between the levels, and the finest level is searched once more in the opposite direction so that vectors which do not lead back lose their confidence. The pyramid grows with the picture, so each level doubles the motion the search can still follow. Its shape follows the optical flow of the AMD FidelityFX SDK (MIT); the passes are this program's own shaders (src/gfx/Shaders.cpp, src/gfx/Pipeline.cpp).

On a Radeon card the FSR host route (src/gfx/FsrHost.cpp) takes the place of the DLSS 5 feature: an FSR 3.1 upscaling context of the FidelityFX API runs at native size with the same colour, depth and motion inputs, and DLSS-NR-on-AMD, a separate program, attaches to that context from the outside and applies the neural pass to its output.

Two threads share the GPU: the processing thread owns the Spout receiver (or the file), the pipeline and a D3D12 queue of its own; the interface thread owns the window, ImGui and a high-priority present queue. Finished pictures are handed over through display buffers with cross-queue fence waits, so the preview always shows the newest completed frame and the window never waits for the neural pass (src/core/App.cpp, src/gfx/Device.cpp). A still image is decoded with WIC (EXIF orientation applied), uploaded once and run through the same pipeline with zero motion for a number of passes until the temporal network settles.

A video file is decoded by a Media Foundation source reader on its own thread (hardware decoder through a DXGI device manager, software fallback) into a short frame queue. The processing thread hands the pipeline one frame at a time and asks for a readback of that frame's result; readbacks are collected in frame order and passed to the writer, so no frame is skipped or duplicated even when the neural pass takes longer than the frame interval. The writer converts each frame to NV12 on a thread of its own and feeds a Media Foundation sink writer (hardware encoder where available) together with the decoded audio samples (src/core/VideoSource.cpp, src/core/VideoWriter.cpp). An animated image goes through the same source interface with its own reader: GIF and APNG frames are decoded by Windows Imaging Component (the APNG chunks are assembled into single PNG frames first) and composited onto the canvas with their disposal and blend rules, animated WebP by libwebp; the writer builds the output animation frame by frame with the source delays rounded cumulatively, so long files keep their timing (src/core/AnimatedImage.cpp).

The Ask AI panel (src/core/AskPanel.cpp) is a WebView2 control that the main window docks at its right edge, created at the first opening only. Its page is the program's own, served under the documentation site's address so Mintlify's widget sees the site it belongs to; it styles the widget with the interface's palette and fonts and talks to the application through web messages (language, theme, notices, links, jumps). The control's window is made the size of the monitor once and clipped to the panel, so resizing the window or the panel moves only the clip and the page never lays itself out again mid-drag. The places an answer can point at are listed in src/ui/JumpTargets.h, each with its names from src/core/I18n.h, its documentation page and where the interface draws it; tools/check_jump_targets.py checks that table against the documentation in every language and against src/ui/MainUI.cpp.

License

MIT (see LICENSE). Third-party components and the NVIDIA notice are listed in THIRD_PARTY_NOTICES.md. This project is not affiliated with VRChat Inc. or NVIDIA Corporation. The DLSS 5 runtime (nvngx_dlssnr.dll) is NVIDIA's software under NVIDIA's terms, not covered by this project's MIT licence; it is used at the user's own risk.

The third-party notices also come in Chinese, Japanese and Korean (docs/THIRD_PARTY_NOTICES.zh-CN.md, .ja.md, .ko.md, shipped next to the executable); the English file is the one that counts. About → Third-party notices opens the one in the interface's language.