Nodes/ComfyUI-H3Studio/MiniMax H3 Temporal LoRA Blend
ComfyUI Node

MiniMax H3 Temporal LoRA Blend

Different LoRA weights before/after a moment IN THE CLIP. Each sampling step runs both models' predictions and blends them along the video timeline with a feathered ramp — one coherent motion trajectory, LoRA influence changing at your chosen second. Wire model_base straight from the checkpoint loader, branch your LoRA loaders off it for model_a (used BEFORE the boundary) and model_b (AFTER; leave empty to drop to the clean base). All three must come from the SAME loaded checkpoint. Costs ~2x sampling time and one extra VRAM copy of the LoRA-touched weights. Needs the model fully resident (no lowvram streaming). Do NOT put the turbo/distill LoRA on only one side — both sides must expect the same sigma schedule.

By shootthesound·Created 13 days ago·Updated 5 days ago· 20
MiniMax H3 Temporal LoRA Blend
  • model_base
  • model_a
  • model_b
  • MODEL
boundary_seconds2.0
feather_seconds0.5
audio_fromramp
Categorymodel/minimax

Inputs (6)

NameTypeDefaultDescription
model_baseMODELThe shared checkpoint, WITHOUT the time-windowed LoRAs. Sampler loads this state.
model_aMODELUsed BEFORE boundary_seconds. Usually model_base -> LoRA loader(s).
boundary_secondsFLOAT2.00–20Clip time where influence hands over from model_a to model_b.
feather_secondsFLOAT0.50–10Width of the linear cross-blend centred on the boundary. 0 = hard switch between adjacent latent frames.
audio_fromCOMBOrampH3 generates audio jointly. 'ramp' blends the audio stream on the same time curve; or pin the whole soundtrack's prediction to one side.
model_boptMODELUsed AFTER boundary_seconds. Empty = the clean base (i.e. the LoRA is dropped at the boundary).

Outputs (1)

NameTypeDescription
MODELMODELWrapped model — feed this to the sampler.