ComfyUI Node
MiniMax H3 Adaptive Cache (Advanced)
Advanced MiniMax H3 adaptive tail-cache controls. This is an approximate, seed-changing acceleration patch with automatic Sol-Attn Morton safety bypass.
MiniMax H3 Adaptive Cache (Advanced)
- model
- model
◄cache_depth0.75►
◄sigma_threshold0.12►
◄window_start0.10►
◄window_end0.90►
◄max_consecutive2►
◄quality_guardaudio_video►
◄content_multiplier1.25►
◄content_ceiling0.75►
◄cache_deviceauto►
◄gpu_safety_mb768►
Categorysampling/custom_sampling/minimax_h3
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| cache_depth | FLOAT | 0.750–0.94 | Fraction of trailing DiT blocks skipped on a cache hit. 0.75 recomputes about 12 of 50 blocks. |
| sigma_threshold | FLOAT | 0.120–1 | Maximum normalized sigma change allowed for a cache candidate. |
| window_start | FLOAT | 0.100–0.49 | — |
| window_end | FLOAT | 0.900.51–1 | — |
| max_consecutive | INT | 20–10 | Maximum consecutive cache hits before a forced full refresh. |
| quality_guard | COMBO | audio_video | 3 options: audio_video, video, off |
| content_multiplier | FLOAT | 1.250.25–4 | A candidate is rejected when its content delta is much larger than the lane's moving average. Lower is safer. |
| content_ceiling | FLOAT | 0.750.05–2 | Hard upper bound for symmetric target-content change. Prevents the first candidate from being trusted blindly. |
| cache_device | COMBO | auto | 3 options: auto, gpu, cpu |
| gpu_safety_mb | INT | 7680–8192 | Free-VRAM reserve used by Auto cache placement. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |