ComfyUI Node
FlashVSR Tiny Decode
Decodes TCDecoder output with four-frame conditioning streamed from CPU, reusable temporal-state buffers, native compute-dtype weights, a true frame-sequential low-VRAM path, and bounded temporal/output batches. Use temporal_batch_size=1 and output_chunk_size=1 for minimum VRAM.
FlashVSR Tiny Decode
- samples
- decoder
- video
- IMAGE
◄output_chunk_size4►
◄temporal_batch_size1►
◄profile_cuda_eventsfalse►
CategoryFlashVSR/decoding
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| samples | LATENT | — | |
| decoder | FLASHVSR_DECODER | — | |
| video | FLASHVSR_VIDEO | — | |
| output_chunk_sizeopt | INT | 41–16 | Decoded frames staged on the GPU before one copy to the final IMAGE tensor. 1 minimizes VRAM; 4 matches TCDecoder's temporal output group. |
| temporal_batch_sizeopt | INT | 11–4 | Latent timesteps decoded together. 1 uses a true depth-first path with one generated high-resolution frame active at a time; 2 often improves GPU utilization substantially; 4 is faster when it fits but can use roughly 1.5-3x the decoder activation VRAM. This remains bounded and does not grow with the complete clip duration. |
| profile_cuda_eventsopt | BOOLEAN | false | Print TCDecoder CUDA-event timings grouped by resolution for conditioning transfer, pixel unshuffle, convolutions, MemBlocks, TGrow, state updates, crop/clamp and output copies, plus wall time and peak allocated/reserved VRAM. Profiling adds a final CUDA synchronization. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |