ComfyUI Node
FlashVSR Sampler (streaming sparse)
A ComfyUI node in FlashVSR with 12 inputs and 1 output.
FlashVSR Sampler (streaming sparse)
- model
- image
- latent
◄width1280►
◄height768►
◄seed0►
◄scale4►
◄kv_ratio3.5►
◄local_range11►
◄sparse_ratio2.0►
◄offloadfalse►
◄chunk_q0►
◄long_modefalse►
CategoryFlashVSR
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| model | FVSRN_MODEL | — | |
| image | IMAGE | — | |
| width | INT | 1280128–16384 | LQ regularised width (nearest-exact resize before x-scale) |
| height | INT | 768128–16384 | LQ regularised height |
| seed | INT | 00–2147483647 | — |
| scale | INT | 41–4 | upscale factor: LQ is bicubic-upscaled x scale, centre-cropped to 128-multiple |
| kv_ratio | FLOAT | 3.50–10 | — |
| local_range | INT | 111–50 | 9 = sharper, 11 = more stable (official recommendation) |
| sparse_ratio | FLOAT | 2.00–10 | — |
| offload | BOOLEAN | false | unload the DiT after sampling |
| chunk_q | INT | 00–8192 | query-chunk size for the sparse attention (0=auto). Must be a multiple of 128. Lower = less VRAM. |
| long_mode | BOOLEAN | false | tiny_long mode: emit per-window latent segments for segmented TCDecoder decoding (long videos); requires tcd_encoder |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | — |