ComfyUI Node
BSAI VedaSparse Patch
A ComfyUI node in BSAI/VedaSparse with 15 inputs and 1 output.
BSAI VedaSparse Patch
- model
- MODEL
◄enabledtrue►
◄keep_percent5►
◄head_tilingdual-fast (2 groups)►
◄custom_tiling4,4,4;8,4,2;2,4,8;4,8,2►
◄tripool_modetriplet (avg+max+min)►
◄scorer_weightsNone (heuristic TripPool)►
◄aspectauto►
◄force_dims►
◄start_percent0.20►
◄end_percent1.00►
◄min_tokens4096►
◄sink_conditioningexact_kv_and_rows►
◄override_priorityauto►
◄verbosefalse►
CategoryBSAI/VedaSparse
Inputs (15)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| enabled | BOOLEAN | true | When off, the model passes through unchanged. |
| keep_percent | FLOAT | 51–100 | Percentage of key tiles to keep per query tile. 10 = 90% sparse (recommended); 5 is too aggressive and may lose structure; 20 is a conservative speed tier. |
| head_tiling | COMBO | dual-fast (2 groups) | Head-Aware Tiling preset. head-aware cycles through 4 spatiotemporal tilings per head (paper-optimal); balanced uses (4,4,4) for all heads; temporal/spatial emphasize time or space. |
| custom_tiling | STRING | 4,4,4;8,4,2;2,4,8;4,8,2 | Tiling table when head_tiling=custom, semicolon-separated 'pt,ph,pw'; each pt*ph*pw must equal 64. |
| tripool_mode | COMBO | triplet (avg+max+min) | Tile descriptor: triplet=Avg+Max+Min (paper-optimal). |
| scorer_weights | COMBO | None (heuristic TripPool) | Distilled predictor weights (optional). Place them in ComfyUI/models/veda_scorers/. No weights = heuristic TripPool scoring. |
| aspect | COMBO | auto | Output aspect ratio. In auto mode, geometry is inferred from the H3 attention pe_index. |
| force_dims | STRING | Manually specify video latent dimensions 'T,H,W' (optional). Usually leave empty. | |
| start_percent | FLOAT | 0.200–1 | Sigma window start (H3 sigma decreases monotonically from 1.0 to 0.0). start=0.5 means the first 50% of steps run dense warm-up (to protect initial structure), then switch to sparse. |
| end_percent | FLOAT | 1.000–1 | Sigma window end (progress percent). end=1.0 means sparse all the way to the end; end=0.95 means the last 5% returns to dense to protect details. |
| min_tokens | INT | 40960–1000000 | If the sequence has fewer tokens than this, run dense (short videos gain little speedup, and it is safer when VRAM is tight). Default 8192. |
| sink_conditioning | COMBO | exact_kv_and_rows | Keep conditioning rows (text/audio) exact. |
| override_priority | COMBO | auto | Cooperation with external attention overrides (official Veda-on-ComfyUI): auto = v3.4 yields when an external override is detected, letting official Veda take over (avoids double-sparsity); veda34 = force v3.4 priority (ignore external overrides); official = always yield. All three modes are non-conflicting and can coexist. |
| verbose | BOOLEAN | false | Print verbose logs. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |