ComfyUI Node
BSAI VDN-H3 Accel · Accel Suite (BlockCache/CacheDiT/TE/VSA)
Acceleration suite switches: Block Cache + CacheDiT + TE-Speed + VSA sparse attention. Injected via model_options; each kernel auto-enables when available and skips silently otherwise. Outputs config_json for the downstream BSAI H3 MotionFix node.
BSAI VDN-H3 Accel · Accel Suite (BlockCache/CacheDiT/TE/VSA)
- model
- MODEL
- config_json
◄block_cache_enabletrue►
◄block_cache_start0.10►
◄block_cache_end0.85►
◄cachedit_enabletrue►
◄te_speed_enabletrue►
◄vsa_enabletrue►
◄vsa_keep_percent20.0►
◄detail_preserve0.00►
CategoryBSAI/VDN-H3
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| block_cache_enable | BOOLEAN | true | Block Cache: cache intermediate block outputs at stride to skip redundant compute (ComfyUI-CacheDiT kernel). |
| block_cache_start | FLOAT | 0.100–1 | Enable block cache from this sampling progress (keep early steps precise). |
| block_cache_end | FLOAT | 0.850–1 | Disable block cache at this progress (keep late steps detailed). |
| cachedit_enable | BOOLEAN | true | CacheDiT diffusion-Transformer cache: skip attention recompute within a t-neighborhood, large speedup. |
| te_speed_enable | BOOLEAN | true | TE-Speed: text-encoder fast path (enable if the matching patch node is present). |
| vsa_enable | BOOLEAN | true | VSA video sparse attention (FastH3 native sparse): keep top-k video blocks under exact attention. |
| vsa_keep_percent | FLOAT | 20.00.5–100 | Percentage of exact attention VSA keeps (higher = more detail, slower). |
| detail_preserve | FLOAT | 0.000–1 | Detail preservation: >0 applies high-frequency compensation for cache/VSA detail loss (small cost, keeps texture). |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| config_json | STRING | — |