ComfyUI Node
Load FlashVSR TCDecoder
A ComfyUI node in FlashVSR/loaders with 5 inputs and 1 output.
Load FlashVSR TCDecoder
- FLASHVSR_DECODER
◄decoder_name▾►
◄compute_dtype▾►
◄fuse_tgrowfalse►
◄channels_lastfalse►
◄compile_memblocksfalse►
CategoryFlashVSR/loaders
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| decoder_name | COMBO | 1 options: No decoder safetensors found in models/flashvsr | |
| compute_dtype | COMBO | 4 options: auto, bf16, fp16, fp32 | |
| fuse_tgrowopt | BOOLEAN | false | Composes each linear temporal-growth 1x1 convolution with its following 3x3 convolution once while loading. This removes three decoder convolutions and their intermediate activations. It is mathematically equivalent, but may select a larger cuDNN workspace. Leave disabled for the low-VRAM baseline. |
| channels_lastopt | BOOLEAN | false | Uses channels-last Conv2d weights and activations. This can improve throughput on some GPUs, but cuDNN may pick algorithms with larger temporary workspaces. Leave disabled for the low-VRAM baseline. |
| compile_memblocksopt | BOOLEAN | false | Experimental: use torch.compile/Inductor only around TCDecoder MemBlocks. The first run is slower and each static block shape may compile separately. Failures automatically return to eager execution. Leave disabled for the compatible Dynamic VRAM baseline. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| FLASHVSR_DECODER | FLASHVSR_DECODER | — |