Nodes/H3 Ref2VA Accelerator/MiniMax H3 Ref2VA Accelerator
ComfyUI Node

MiniMax H3 Ref2VA Accelerator

Production-oriented MiniMax H3 Ref2VA accelerator. Balanced is the recommended quality/speed profile. Block 0 runs every step; blocks 1-49 are reused only when global, target-video, target-audio, reference and temporal guards all pass. This is deterministic but approximate. CPU cache storage is recommended for offloaded pruned BF16 H3.

By BMB12d3·Created 22 days ago·Updated 20 days ago· 8
MiniMax H3 Ref2VA Accelerator
  • model
  • MODEL
modeRef2VA Balanced
cache_storageCPU (VRAM-safe)
debugfalse
tail_rescalefalse
global_threshold0.090
video_threshold0.090
audio_threshold0.080
visual_ref_threshold0.065
audio_ref_threshold0.065
temporal_threshold0.110
start_percent0.10
end_percent0.95
max_consecutive_hits1
cpu_tail_computeSafe CPU (v0.3 behavior)
CategoryMiniMax H3/optimization

Inputs (15)

NameTypeDefaultDescription
modelMODEL
modeCOMBORef2VA Balanced6 options: Ref2VA Balanced, Ref2VA Conservative, Ref2VA Ultra Safe (quality-first), Ref2VA Aggressive (experimental), Observe Only (no caching), Custom
cache_storageCOMBOCPU (VRAM-safe)CPU is recommended for pruned BF16/offloaded H3. GPU can be slightly faster for smaller quantized checkpoints with comfortable VRAM headroom.
debugBOOLEANfalse
tail_rescaleoptBOOLEANfalseExperimental only. Fixed-seed testing produced slightly different output without a clear quality benefit. Production recommendation: leave OFF.
global_thresholdoptFLOAT0.0900–1Custom mode only: overall block-0 residual-change limit.
video_thresholdoptFLOAT0.0900–1Custom mode only: target-video residual-change limit.
audio_thresholdoptFLOAT0.0800–1Custom mode only: target-audio residual-change limit.
visual_ref_thresholdoptFLOAT0.0650–1Custom mode only: visual-reference residual-change limit.
audio_ref_thresholdoptFLOAT0.0650–1Custom mode only: reference-audio residual-change limit.
temporal_thresholdoptFLOAT0.1100–1Custom mode only: worst target-video frame-change limit.
start_percentoptFLOAT0.100–1Custom mode only: earliest denoising fraction where cache hits are allowed.
end_percentoptFLOAT0.950–1Custom mode only: latest denoising fraction where cache hits are allowed.
max_consecutive_hitsoptINT11–10Custom mode only. Quality-first recommendation: leave at 1.
cpu_tail_computeoptCOMBOSafe CPU (v0.3 behavior)CPU cache mode only. Production recommendation: Safe CPU. Auto GPU Fast Path is retained for benchmarking; on the validated RTX 5090 BF16 workflow it saved only ~2 seconds over an ~11 minute sampler run.

Outputs (1)

NameTypeDescription
MODELMODEL