Nodes/ComfyCollectorNodes/MiniMax H3 Ref Tinker (CCN)
ComfyUI Node

MiniMax H3 Ref Tinker (CCN)

Experimental control over MiniMax H3 reference conditioning: reference strength (cond-row noise aug) and reference RoPE geometry (temporal gap to the target streams, spatial offset of image refs). Wraps the diffusion-model call; composes with other model patches.

By valkymaera·Created 3 months ago·Updated 4 days ago· 1
MiniMax H3 Ref Tinker (CCN)
  • model
  • MODEL
enabledtrue
visual_cond_noise_aug0.999
audio_cond_noise_aug1.000
ref_temporal_gap0.0
ref_spatial_shift_h0.00
ref_spatial_shift_w0.00
Categorymodel/patch/minimax

Inputs (7)

NameTypeDefaultDescription
modelMODEL
enabledBOOLEANtrueManual off switch - passes the model through unpatched.
visual_cond_noise_augFLOAT0.9990–1Strength of visual condition rows (reference images and fl2va keyframes). 0.999 = stock (rows re-injected ~clean every step). Lower mixes fixed-seed noise into the rows (aug*latent + (1-aug)*noise) and lowers their declared timestep, weakening verbatim copying before identity. Useful band ~0.90-1.00.
audio_cond_noise_augFLOAT1.0000–1Same for reference-audio rows. 1.0 = stock (fully clean).
ref_temporal_gapFLOAT0.0-1–2000Extra RoPE-time distance between reference blocks and the target audio/video streams, in position ticks (1 tick = 1/40 s on the shared AV clock). Stock ref->video gap is 1.0 tick; video frame hops are 1.67/6.67. -1.0 co-locates the last image ref with frame 0 (soft keyframe-anchor mode). 0 = stock.
ref_spatial_shift_hFLOAT0.00-64–64Offset added to image-reference RoPE h coordinates (frame grids are area-normalized, roughly spanning 0-32). Decorrelates ref<->target pixel alignment without editing the image. 0 = stock.
ref_spatial_shift_wFLOAT0.00-64–64Offset added to image-reference RoPE w coordinates. 0 = stock.

Outputs (1)

NameTypeDescription
MODELMODEL