ComfyUI Node
RTX Frame Generation
RTX Frame Generation: NVIDIA pair interpolation, RGB only (alpha discarded). Multiplier keeps every source frame and endpoints: (N-1)*M+1 frames. Encode at source FPS*M. Specific timestep inserts one frame per pair; positions are nonuniform unless timestep=0.5. One complete IMAGE output is retained.
RTX Frame Generation
- images
- images
◄generation_type▾►
◄multiplier2►
◄timestep0.50►
◄modeHIGH►
◄automatic_shot_change_detectiontrue►
◄shot_changefalse►
◄device_id0►
◄image_encoding8-bit RGB►
◄auto_unload_modelstrue►
◄use_mmapfalse►
CategoryDaSiWa/Video
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| images | IMAGE | Ordered IMAGE batch: still images or decoded video frames, not a VIDEO object. RGB only; alpha discarded. No FPS, timestamps or audio are passed through. | |
| generation_type | COMBO | frame rate multiplier: evenly insert M-1 frames per pair, output (N-1)*M+1; encode at source FPS*M. specific timestep: insert one frame per pair, output 2*N-1; only 0.5 is uniformly spaced (encode at source FPS*2). Both preserve endpoints; container duration can differ by a final frame interval. | |
| multiplier | INT | 22–16 | Active only for frame rate multiplier: M subdivisions per source interval, M-1 generated frames per pair. Output (N-1)*M+1; set encoder FPS to source FPS*M. Ignored for specific timestep. |
| timestep | FLOAT | 0.500.01–0.99 | Active only for specific timestep: fractional position between previous (0) and next (1) source frame. 0.5 is the midpoint; other values produce unequal temporal spacing that a constant-FPS IMAGE encoder cannot reproduce exactly. Ignored for multiplier mode. |
| mode | COMBO | HIGH | SDK interpolation quality preset: LOW favors speed, MEDIUM is the middle tier, HIGH favors quality. These are frame-generation modes, not VSR levels; cost and visual quality depend on SDK, GPU and content, with no fixed speed or quality guarantee. |
| automatic_shot_change_detection | BOOLEAN | true | On: ask the SDK to detect cuts automatically between source frames. Off: disable automatic detection; the separate shot_change flag still applies. Cut handling is SDK-defined, not a per-frame cut list. |
| shot_change | BOOLEAN | false | On: explicitly mark every submitted source pair as a cut (useful for a single pair). Off: do not force cuts; automatic detection can still apply. This is not a list of cut indices. |
| device_id | INT | 00–8 | CUDA device index for NVIDIA interpolation (0 is the first GPU). Must identify an available supported NVIDIA device; does not choose encoder or audio device. |
| image_encodingopt | COMBO | 8-bit RGB | 8-bit RGB: default SDK encoding, retains source floating output dtype. 10-bit RGB (RGB10A2): packed 10-bit SDR SDK processing, float32 output; not HDR or smaller IMAGE storage. Original frames are copied directly. |
| auto_unload_modelsopt | BOOLEAN | true | On: reclaim targeted GPU memory and, under RAM output pressure, native execution caches and model pins; re-check before allocation. Off: no reclamation. Never resets the running executor or globally offloads GPU models into RAM. Native cache eviction requires ComfyUI RAM-pressure cache support; full batches still need storage. |
| use_mmapopt | BOOLEAN | false | Off: never allocate disk output. On: allow a temporary mmap as the last output tier if budgeted VRAM/RAM cannot fit the complete IMAGE batch. Requires disk space for the full output; not streaming input/output. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| images | IMAGE | — |