DLSS5 Basic Workflow

| Input | Type | Notes |
| --- | --- | --- |
| `image` | IMAGE | Batch = temporal order |
| `video` | VIDEO | From LoadVideo; frames processed in playback order |
| `upscaling_mode` | combo | `1x (DLAA)` native enhance, `1.5x / 1.724x / 2x / 3x` fixed NVIDIA factors (carrier + feature 18) |
| `style` | combo | default / natural / cinematic |
| `preset` | INT 0–9, default 0 | Internal render preset hint, passed to `DLSSNR.Hint.Render.Preset`. Keep 0 unless the docs of your runtime build define another value |
| `intensity` | FLOAT 0.0–2.0, default 1.0 | Overall neural pass strength (`DLSSNR.Intensity`). 1.0 is full strength; >1.0 usually has no extra effect; <1.0 blends back towards the source for lighter denoising |
| `tone` | FLOAT 0.0–2.0, default 1.0 | Local tone mapping strength (`DLSSNR.LocalToneStrength`; the global term `GlobalToneStrength` is fixed at 1.0). Higher = punchier local contrast, lower = flatter |
| `structure` | FLOAT 0.0–2.0, default 1.5 | Local detail / structure reconstruction strength (`DLSSNR.LocalStructureStrength`). Higher = sharper with more texture; lower = softer and cleaner (drop it if the result looks plasticky) |
| `skin` | FLOAT −1.0–2.0, default 2.0 | Skin structure reconstruction strength (`DLSSNR.SkinStructureStrength`). **Only active with `auto_mask=on`**; −1 leaves the decision to the model |
| `auto_mask` | combo on/off, default on | Automatic skin-region detection (`DLSSNR.UseAutoMask`): the model finds skin itself and applies `skin` there — effectively the on/off switch for `skin` |
| `motion` | combo | auto (optical flow when OpenCV present) / optical_flow / none |
| `scene_change_threshold` | FLOAT | Scene-cut history reset threshold (optical flow only) |
| `batch_mode` | combo | `temporal sequence` keeps history between frames; `still images` resets every frame |
| `warmup_frames` | INT | Warm-up budget reported to the worker (0 is fine) |
| `keep_audio` | combo | Keep original audio in the VIDEO output when supported |
| `backend` | combo | auto / linux-wine / windows-bridge |
| `channel_order` | combo | auto-detects R/B swap of the runtime output; force RGBA/BGRA on wrong colours |
| `runtime_dir` | STRING | Optional runtime folder override; when empty the lookup order is `DLSS5_RUNTIME_DIR` → `<ComfyUI>/models/dlss5` → the bundled `runtime/` |
| `wine_prefix` | STRING | Linux: WINEPREFIX override for the worker |

| Output | Type | Meaning |
| --- | --- | --- |
| `images` | IMAGE | Processed frames (set when `image` was connected) |
| `video` | VIDEO | Processed video object for SaveVideo (set when `video` was connected; on old ComfyUI builds without the VIDEO type this is None and the `status` output names the written file) |
| `status` | STRING | One-line summary: frames, sizes, backend, channel order, output notes |