ComfyUI-RH-DLSS5 Introduction
ComfyUI-RH-DLSS5 brings NVIDIA DLSS Neural Rendering, optional upscaling, and frame generation into your ComfyUI workflows. You can use it to refine images, process video frames, increase resolution, or create smoother motion by adding AI-generated frames between existing ones.
The extension provides two nodes:
| Node | Main purpose | Where to find it |
|---|---|---|
RH DLSS5 Enhance (NR / Upscale) — RH_DLSS5Enhance | Apply neural enhancement at the original resolution or enlarge images and video frames | image/upscaling |
RH DLSS5 Frame Interpolation — RH_DLSS5FrameInterpolation | Increase a sequence’s frame rate using NVIDIA DLSS frame generation | video |
What you can use it for
- Refine AI-generated artwork: adjust local contrast and reconstructed detail, including skin and material structure.
- Clean up generated video: use neural enhancement to reduce noise or compression artifacts while controlling how much texture remains.
- Create larger deliverables: choose a fixed enlargement factor or a resolution target such as 2K or 4K.
- Make animation look smoother: turn a 24 fps clip into a 48 fps clip with
2xframe interpolation, for example. - Keep processing inside ComfyUI: use standard
IMAGEandVIDEOconnections rather than a separate online service. - Preserve sound where supported: retain source audio or provide replacement audio to the frame-interpolation node.
The extension calls NVIDIA’s native NGX runtime. It does not imitate DLSS with a look-alike filter, and it does not send your frames to an HTTP processing service.
Compatibility at a glance
- GPU: a compatible NVIDIA RTX GPU and NVIDIA runtime build are required. Compatibility depends on the combination of GPU, driver, and runtime—not just the RTX name.
- Windows: Windows 10/11 x64. The enhancement node uses an in-process D3D12 bridge.
- Linux: uses Windows native workers through Wine, together with DXVK, vkd3d-proton, and DXVK-NVAPI.
- ComfyUI: native video workflows need a build with
VIDEO,LoadVideo, andSaveVideosupport. Image workflows use the standardIMAGEtype.
The author reports end-to-end Linux testing on an RTX 4090. See the models and verification sections for the distinction between tested configurations and runtime-dependent support.
How ComfyUI-RH-DLSS5 Works
The two nodes solve different problems: enhancement changes the appearance or size of frames; interpolation adds frames to the timeline.
Neural enhancement: refining what is already in the frame
Think of enhancement as a finishing pass. You provide an image, and NVIDIA’s neural renderer reconstructs its appearance using the available visual information.
In a game, neural rendering can receive additional information about geometry, lighting, and materials. Ordinary images and videos do not contain those engine buffers. Here, the renderer works from your frames and motion guidance estimated from the footage.
That distinction matters: this is not the same as supplying a complete 3D scene. You should judge the result as a reconstruction based on the source, not as a guaranteed recovery of the original physical lighting or missing details.
The enhancement pipeline is:
Image or video frames
→ Frame preparation and motion guidance
→ Optional DLSS Super Resolution enlargement
→ DLSS Neural Rendering
→ Processed IMAGE or VIDEO
- At 1x, neural enhancement runs at the original dimensions.
- Above 1x, ordinary DLSS Super Resolution first supplies the enlarged image. Neural Rendering then processes that higher-resolution surface.
For example, a 480 × 736 portrait processed at 2x becomes 960 × 1472.
Motion guidance: telling the renderer what moved
Optical flow estimates how visible objects move between neighboring frames. Imagine drawing small arrows over the picture: each arrow indicates where a patch of the image moved.
These arrows help processing follow a coherent sequence rather than treating every frame as unrelated.
- For a still image, zero motion is appropriate.
- For a video sequence, estimated motion provides temporal guidance.
- For unrelated images in one batch, use
still imagesso each image starts fresh.
Scene cuts: clearing the visual memory
When a clip cuts from one shot to another, information from the previous shot should not influence the new one. The extension resets its temporal history at detected scene changes.
The enhancement node also treats implausible flow—such as unreliable estimates during fast movement or noisy footage—as a reason to reset history, helping prevent accumulated ghosting.
Frame interpolation: drawing the moments between frames
Interpolation creates new images between existing frames rather than blending two pictures together.
At 2x, a sequence can be thought of like this:
Original: A B C
Interpolated: A AB B BC C
AB and BC are generated intermediate frames.
For a continuous sequence with T source frames, the documented 2x path produces 2T − 1 frames at exactly twice the source frame rate. It does not invent an additional interval beyond the last source frame. Generated frames are skipped across detected scene cuts.
The 3x and 4x modes use cascaded processing, so they require more work than 2x.
Streaming: why VIDEO inputs matter for longer clips
For frame interpolation, a VIDEO input follows a streaming path: output frames go directly into the video encoder instead of accumulating as one full-length floating-point image batch.
This keeps host-memory use bounded with respect to clip length, though actual memory requirements still depend on resolution, decoding, and the chosen multiplier.
An IMAGE batch returns all output images in memory. That is useful for further image processing, but it can be much more demanding for long sequences.
ComfyUI-RH-DLSS5 Features
Enhance images and videos with one node
RH_DLSS5Enhance accepts image, video, or both.
| Connection | What it provides |
|---|---|
image input | A still image or an ordered batch of images |
video input | A native ComfyUI video, processed in playback order |
images output | Processed images when an image input was supplied |
video output | Processed video when a video input was supplied |
status output | A short summary of frame counts, dimensions, backend, channel order, and output notes |
Connect images to SaveImage and video to SaveVideo.
Choose native enhancement or upscaling
upscaling_mode | What it does | Example use |
|---|---|---|
1x (DLAA) | Enhances without increasing dimensions | Finish an image already at delivery size |
1.5x | Multiplies width and height by 1.5 | A moderate enlargement |
1.724x | Uses NVIDIA’s balanced scaling factor | An intermediate choice between 1.5x and 2x |
2x | Doubles width and height | Turn 960 × 540 into 1920 × 1080 |
3x | Triples width and height | A larger enlargement from a smaller source |
1K / 2K / 4K / 8K | Automatically resolves the scale needed for a target resolution bucket | Choose a delivery target rather than a manual factor |
The automatic buckets use these short-edge targets:
| Bucket | Target short edge | Documented 16:9 output |
|---|---|---|
1K | 1080 pixels | 1920 × 1080 |
2K | 1440 pixels | 2560 × 1440 |
4K | 2160 pixels | 3840 × 2160 |
8K | 4320 pixels | 7680 × 4320 |
The extension preserves aspect ratio and selects the smallest supported factor that reaches the target. An input already at or above the bucket stays at 1x; the node does not downscale it.
A request fails with an explanatory error if the target cannot be reached within 3x or would exceed the documented 7680 × 4320 output envelope.
A scaling factor applies to both dimensions. A
3xenlargement contains approximately nine times as many pixels per frame, which can substantially increase processing and memory requirements.
Control the enhancement look
These controls map to NVIDIA’s DLSS Neural Rendering properties. Their exact effect can vary with the supplied runtime.
| Control | Range or choices | How to use it |
|---|---|---|
style | default, natural, cinematic, off (bypass NR) | Select an overall style profile, or bypass all processing |
intensity | 0.0–2.0; default 1.0 | Overall neural-pass strength. Lower values blend toward the source; values above 1.0 usually add no further effect |
tone | 0.0–2.0; default 1.0 | Higher values increase local contrast; lower values produce a flatter appearance |
structure | 0.0–2.0; default 1.5 | Higher values emphasize reconstructed detail and texture; lower values give a softer, cleaner result |
skin | −1.0–2.0; default 2.0 | Skin-specific structure strength; −1 leaves the decision to the model |
auto_mask | on / off; default on | Lets the runtime identify skin regions and apply skin there |
preset | Integer 0–9; default 0 | Internal rendering hint. Keep 0 unless your runtime documentation defines another value |
Important: skin only has an effect when auto_mask=on.
style=off is a complete bypass, not an “upscale without enhancement” mode. Image batches are returned unchanged, and file-backed videos pass through without decoding or re-encoding.
Suggested starting points
| Your goal | Starting settings | What to inspect |
|---|---|---|
| Clean up generated video | intensity=1.0, default structure | Raise structure if detail disappears; lower it if noise remains |
| Refine a portrait | auto_mask=on, skin=2.0 | Check skin texture alongside hair, clothing, and background detail |
| Reduce an over-processed look | Lower structure first | If still too strong, reduce intensity |
| Add local contrast | Increase tone gradually | Watch shadow detail and bright areas |
| Finish at the existing size | 1x (DLAA) | Compare texture and tone without a resolution change |
Change one setting at a time on the same image or short clip. This makes it easier to tell whether an improvement comes from structure, tone, or overall intensity.
Manage motion and frame relationships
| Control | Options | Purpose |
|---|---|---|
motion | auto, optical_flow, none | auto uses optical flow when OpenCV is available; none supplies zero motion |
scene_change_threshold | Floating-point threshold | Controls optical-flow-based scene-cut history resets |
batch_mode | temporal sequence, still images | Keep history for a coherent sequence, or reset it for every independent image |
warmup_frames | Integer | Reports a warm-up budget to the worker; 0 is a valid starting point |
keep_audio | On/off | Retains original audio in video output where supported |
Keep video frames in playback order. A batch is interpreted as a sequence, not as a collection to be automatically sorted.
Increase frame rate with DLSS frame generation
RH_DLSS5FrameInterpolation accepts either a video or an ordered image batch.
| Control | Options or behavior |
|---|---|
output_fps | 1x, 2x, 3x, 4x, or exact target-rate presets spanning 23.976–144 fps |
motion | auto uses hardware NVOFA when available, otherwise OpenCV DIS; nvof or dis forces a guide |
scene_change_threshold | 0.01–1.0, default 0.24; higher values cause fewer scene-cut resets |
keep_audio | Default on; off removes audio during processing |
audio | Optional replacement audio for video, or an audio source for image-batch inputs |
For exact target rates, the node builds the smallest suitable 2x / 4x / 8x dense grid and selects frames for the requested timeline without duplicating frames. The target must be higher than the source rate and no more than 6x the source rate.
Examples:
- 24 → 48 fps: choose
2x, the author’s recommended starting point. - 24 → 72 fps: choose
3x. - 24 → 60 fps: choose the exact 60 fps target.
- No interpolation: choose
1x.
1x is a true passthrough: it does not interpolate or re-encode. It ignores motion and audio-processing options and keeps the source audio.
Understand interpolation inputs and outputs
| Input type | Timing and audio | Outputs |
|---|---|---|
VIDEO | Uses source frame rate and source audio, unless replacement audio is supplied | Streaming-encoded video; images is None |
IMAGE batch | Assumes a 24 fps source; attach sound through audio | Full images batch plus an in-memory video object |
When possible, source-file audio is stream-copied without re-encoding. Replacement audio or audio without a usable source file is encoded to AAC when needed.
Advanced runtime and backend controls
Most workflows can leave these controls at their defaults.
| Control | Purpose |
|---|---|
backend | Enhancement node: auto, linux-wine, or windows-bridge |
channel_order | Enhancement node: automatic color-order detection, with manual RGBA or BGRA overrides |
runtime_dir | Overrides the runtime folder for the selected node |
wine_prefix | Overrides the Wine environment used on Linux |
The enhancement runtime lookup is:
runtime_dir
→ DLSS5_RUNTIME_DIR
→ ComfyUI/models/dlss5
→ extension runtime/ fallback
The interpolation runtime lookup is:
runtime_dir
→ DLSS5_FG_RUNTIME_DIR
→ ComfyUI/models/dlss5/dlssg
→ extension runtime/dlssg fallback
For managed or multi-GPU systems, the extension resolves CUDA_VISIBLE_DEVICES and attempts to keep Wine workers on the assigned GPU. Advanced overrides include DLSS5_GPU_INDEX, DLSS5_WINE, and DLSS5_WINEPREFIX. These are troubleshooting controls rather than artistic settings.
Queueing and caching
Each run starts with fresh processing state, so temporal history is not intended to leak from one queued job into another.
- Linux launches a Wine worker for each execution, with execution serialized by an internal lock.
- Windows enhancement reuses a bridge session on the same GPU when dimensions and settings match.
- Identical work may be served from ComfyUI’s cache, based on settings and input probes or video metadata.
Re-queueing the same workflow therefore does not necessarily start another render.
ComfyUI-RH-DLSS5 Models and Runtime Components
This extension does not offer a collection of artist-selectable checkpoint models. It uses NVIDIA runtime libraries, with optional GPU-specific Neural Rendering builds selected automatically.
What each component does
| Component | Role | When it is needed |
|---|---|---|
nvngx_dlssnr.dll | DLSS Neural Rendering, NGX feature 18 | Required for enhancement |
nvngx_dlssnr_rtx30.dll | Optional RTX 30-lineage Neural Rendering build | Preferred automatically on RTX 30 GPUs when present |
nvngx_dlssnr_rtx40.dll | Optional RTX 40-lineage Neural Rendering build | Preferred automatically on RTX 40 GPUs when present |
nvngx_dlss.dll | DLSS Super Resolution carrier | Enhancement modes above 1x |
_nvngx.dll | NVIDIA NGX core loader | Runtime infrastructure |
nvngx_dlssg.dll | DLSS frame generation | Frame interpolation |
dlssg-worker.exe | Native host that drives frame generation | Frame interpolation |
nvngx.dll in dlssg/ | NGX SDK loader stub | Frame interpolation |
These files are not interchangeable. For example, renaming nvngx_dlss.dll to _nvngx.dll does not replace the NGX loader.
Automatic GPU-specific selection
| GPU family | Selection order | Verification reported by the author |
|---|---|---|
| RTX 30 / Ampere | RTX 30 build → RTX 40 build → universal build | Not yet tested |
| RTX 40 / Ada | RTX 40 build → universal build | RTX 4090 tested |
| RTX 50 / RTX 6000D, SM120 | Universal build | RTX 6000D tested |
The documentation describes a universal build covering RTX 20 through RTX 50, but does not establish verified RTX 20 operation. Stock SDK builds may have narrower support, including RTX 50-only builds. Treat support as specific to the runtime you are authorized to use.
For advanced diagnostics, DLSS5NR_SNR_FILENAME can override automatic selection with a plain runtime filename.
What not to copy from related extensions
Related projects expose controls such as model presets J / K / L / M, dedicated encoding options, or separate settings nodes. Those are not documented controls of ComfyUI-RH-DLSS5.
Likewise, runtime-download scripts and bundled-runtime behavior described by other projects should not be assumed to apply here. This extension does not download or redistribute NVIDIA runtime libraries or the frame-generation worker.
Use only runtime files you are legally entitled to use. See runtime/README.txt for component details and the official NVIDIA DLSS releases for publicly available SDK releases.
ComfyUI-RH-DLSS5 Workflows and Examples
Enhance a still image
LoadImage
→ RH DLSS5 Enhance (NR / Upscale)
→ SaveImage
- Connect the image to
image. - Start with
1x (DLAA)to evaluate enhancement separately from enlargement. - Use
still imagesfor independent artwork. - Adjust structure, skin, or intensity after comparing against the source.
- Select an upscaling factor if you need a larger result.
Example: image API workflow.
Enhance a video
LoadVideo
→ RH DLSS5 Enhance (NR / Upscale)
→ SaveVideo
- Connect the source to
video. - Use
temporal sequence. - Start with automatic motion guidance.
- Keep audio enabled if the result should retain sound.
- Connect the
videooutput to SaveVideo.
Example: video enhancement API workflow.
Make motion smoother
LoadVideo
→ RH DLSS5 Frame Interpolation
→ SaveVideo
- Begin with
output_fps=2x. - Leave
motion=auto. - Review fast motion, object boundaries, and scene transitions.
- Move to
3x,4x, or an exact target only when the additional smoothness is useful.
Example: frame-interpolation API workflow.
The supplied examples are API-format JSON workflows. Replace their placeholder media names before submitting them through a compatible workflow interface.
ComfyUI-RH-DLSS5 Current Improvements and Verification
The supplied documentation does not include a numbered release history. The following are documented improvements and checks, rather than an invented version-by-version changelog.
Streaming frame interpolation
VIDEO interpolation now sends generated frames directly to an FFmpeg video encoder instead of constructing a full-length floating-point output batch.
Why this matters:
- Longer clips no longer require output memory to grow as one complete floating-point sequence.
2xavoids an intermediate generated-frame grid buffer.3x / 4xretain one compact 8-bit grid buffer between cascade stages.- Original audio can be stream-copied from the source.
The author reports a 400-frame, 1080p, 2x test completing at roughly 20 GB peak host RAM. The earlier non-streaming implementation was terminated by the memory limit on the same clip. This is a specific measurement, not a general memory guarantee.
Linux execution and recovery handling
Documented behavior includes:
- Automatic management of a shared virtual display on headless systems.
- Handling of stale display locks left by interrupted workers.
- Wine-prefix validation with actionable errors.
- GPU selection intended to respect scheduler-assigned devices.
These changes help reduce failures on servers and shared multi-GPU systems.
Reported verification
The author reports:
- Enhancement at
1xand1.5xon a 48-frame,320 × 240test clip. - Linux testing with Ubuntu 22.04, Wine 11.0, DXVK 3.1, vkd3d-proton 3.0.1, and DXVK-NVAPI.
- RTX 4090 testing with driver 580.95.05 and DLSSNR 310.8.
- A fresh Ubuntu 22.04 container validation in September 2026.
- Frame-interpolation checks for
2x / 3x / 4x, source-frame alignment, NVOFA and DIS guidance, and audio stream-copy.
The Windows enhancement bridge follows the same protocol as the referenced native integration; the detailed end-to-end verification above was reported for Linux.
Troubleshooting ComfyUI-RH-DLSS5
Start with a small, simple test
Before changing advanced runtime settings:
- Use one image or a short clip.
- Choose
1x (DLAA)for enhancement or2xfor interpolation. - Leave backend and channel order on
auto. - Read the node’s
statusoutput. - Inspect ComfyUI’s console for the complete error message.
Record the GPU, driver, selected runtime, dimensions, and settings when comparing failures.
The result looks too soft, artificial, or over-processed
- Compare at the same display size.
- If detail is disappearing, raise
structuregradually. - If the result looks plasticky or overly reconstructed, lower
structure. - If the overall effect is still too strong, lower
intensity. - For portraits, confirm
auto_mask=onbefore evaluatingskin.
Increasing intensity above 1.0 is usually not a way to recover more detail.
Skin settings appear to do nothing
- Turn
auto_maskon. - Confirm that the image contains visible skin regions.
- Compare two skin-strength settings on the same source.
- Keep other settings unchanged during the comparison.
With automatic masking off, the documented skin-strength control is inactive.
Red and blue appear swapped
- Leave
channel_order=autofor the initial test. - If colors remain wrong, try
RGBA. - If necessary, try
BGRA. - Compare against the source and keep the correct setting.
This override handles runtime builds that expose different red/blue channel ordering.
The requested resolution fails
- Check the input dimensions and selected factor.
- Try a lower fixed factor.
- For an automatic bucket, check whether reaching its short edge would require more than
3x. - Check whether the output would exceed the documented size envelope.
An automatic bucket does not downscale an oversized source.
Video shows ghosting or influence from another image
- Confirm that frames are in playback order.
- Use
temporal sequenceonly for one coherent sequence. - Use
still imagesfor unrelated artwork in a batch. - Check motion guidance and scene-cut settings.
- For interpolation, lower the scene-change threshold if real cuts are not being detected; higher thresholds produce fewer resets.
A runtime file cannot be found
- Read the error to identify the missing component.
- Check the selected node’s
runtime_dir. - Check whether an environment override points to an old folder.
- Verify that the required files are present and correctly named.
- Consult runtime/README.txt.
Enhancement and interpolation use different runtime folders. A working enhancement runtime does not by itself provide frame-generation support.
Feature creation fails with 0xBAD00001
This can indicate that the runtime rejects the selected GPU/driver combination. On headless Linux systems, it can also result from GPU identification failure.
- Confirm that the selected runtime supports your GPU.
- Check the NVIDIA driver.
- On Linux, look for
Failed to get NvPhysicalGpuHandle for LUID. - If that message appears on a machine without an active display output, review DXVK-NVAPI compatibility.
The documented headless issue involves DXVK-NVAPI versions up to 0.9.2 and an invalid Vulkan device LUID. A build with a DXGI adapter-LUID fallback, or a suitable virtual display/Xorg configuration with a connected output, is the documented remedy.
Interpolation returns no images
For a VIDEO input, this is expected.
- Connect the node’s
videooutput to SaveVideo. - Use an
IMAGEbatch input only if you need the entire interpolated image batch in memory.
Audio is missing or was not replaced
- Check
keep_audio. - Confirm that the source video contains audio.
- For image batches, connect an
AUDIOinput. - For replacement sound, connect
audioand leavekeep_audioon. - Check whether
output_fps=1x: passthrough keeps the original video unchanged and ignores audio-processing options.
Memory use is too high
- Prefer the
VIDEOinput for long interpolation jobs. - Try
2xbefore3xor4x. - Test at a smaller resolution.
- Avoid returning a long sequence as a full
IMAGEbatch unless necessary.
Streaming reduces accumulation of output frames; it does not eliminate decoding buffers or other memory costs.
Linux-specific failures
| Symptom | What to check next |
|---|---|
failed to load l_intl.nls during Wine initialization | Wine’s shared data files are incomplete; have the environment maintainer check the Wine installation |
Only llvmpipe appears | Without DISPLAY, this can be normal. Verify NVIDIA visibility with a valid display environment |
ERROR_OUT_OF_HOST_MEMORY from vulkaninfo | Check for a broken or empty nvidia_icd.json; Vulkan loader diagnostics can identify a JSON parsing failure |
vr -3 or VK_ERROR_INITIALIZATION_FAILED | Check for incomplete NVIDIA driver libraries, especially libnvidia-rtcore in containers |
Could not create a D3D12 device | Check Vulkan operation and DISPLAY before investigating the node |
| A run hangs and subsequent retries also hang | An orphaned Wine worker or server may still hold GPU resources or pipes |
Do not wrap ComfyUI’s worker execution in xvfb-run. The author documents that it can leak logs into the binary frame stream; the node manages its own virtual display.
For a hung Linux environment, an experienced operator can stop leftover processes before retrying:
pkill -f 'dlss5nr_hos[t]'
pkill -x wineserver
pkill -x Xvfb
Caution: these commands can affect other Wine or virtual-display workloads under the same user. On shared systems, ask the environment maintainer to isolate and stop only the affected processes.
Frequently asked questions
Does interpolation make a clip play in slow motion?
The multiplier modes increase the output frame rate along with the frame count. They are intended to improve smoothness, not automatically turn the clip into slow motion.
Why does 2x not always mean exactly twice as many frames?
The documented continuous-sequence count is 2T − 1: there is no future source frame after the last frame. Scene cuts also skip generated frames for the affected intervals.
Can I process unrelated images together?
Yes. Use the enhancement node’s still images batch mode to reset history for every image.
Does style=off keep upscaling active?
No. It bypasses all processing.
Can I use an exact target frame rate below the source rate?
No. Exact interpolation targets must exceed the source rate and remain within 6x.
Why did re-queueing not start processing again?
ComfyUI may reuse a cached result for identical settings and inputs. Change a relevant input or setting when you actually need a new evaluation.
Does the extension save the final result automatically?
Use SaveImage or SaveVideo in your workflow to save the output. The interpolation VIDEO path uses a streaming-encoded temporary video.
Learn More about ComfyUI-RH-DLSS5
Practical starting resources
- Image enhancement example — a minimal LoadImage → Enhance → SaveImage workflow.
- Video enhancement example — native VIDEO processing.
- Frame-interpolation example — a starting point for
2xinterpolation. - Runtime component notes — useful when diagnosing missing or incompatible runtime files.
NVIDIA documentation
- NVIDIA DLSS repository — official SDK information.
- NVIDIA DLSS releases — public SDK release packages.
- NVIDIA Optical Flow programming guide (https://docs.nvidia.com/video-technologies/optical-flow-sdk/nvofa-programming-guide/index.html) — optional technical background on motion guidance.
Related implementations
These projects provide implementation background. Their node names, runtime packaging, settings, and troubleshooting procedures are not necessarily interchangeable with this extension.
- ComfyUI-DLSS5-NR-Linux — Linux/Wine reference.
- ComfyUI-DLSS5-NR — Windows native-bridge reference.
- ComfyUI-DLSS5-Enhancer — video-worker protocol reference.
- ComfyUI-NVIDIA-DLSS-Frame-Interpolation — frame-generation worker and scheduling reference.
Linux compatibility resources
Use these when the failure concerns graphics translation, Wine, or GPU discovery rather than an artistic setting.
ComfyUI-RH-DLSS5 License and Attribution
The extension’s Python code and project-owned native components are licensed under MIT. See LICENSE.
The native bridge and host derive from the MIT-licensed reference integrations linked above. The frame-generation pipeline also draws on the referenced worker protocol and scheduling designs.
NVIDIA NGX/DLSS runtime binaries are proprietary NVIDIA software, not covered by the extension’s MIT license. They are not included, downloaded, or redistributed by this plugin. The frame-generation worker is also user-supplied.
Use only components you are authorized to use and follow their applicable license terms. DLSS and NGX are NVIDIA trademarks.
