SanDiegoDude/H3-Turbo-6-Step-LoRA-Comfy

MiniMax-H3 Turbo LoRA — 6-step, native ComfyUI (BF16)

ComfyUI-native conversions of larryvrh/MiniMax-H3-Turbo-Lora, so they load with the stock LoraLoader against core MiniMax-H3 support — no custom nodes, no custom sampler.

All credit for the actual training goes to larryvrh. These are just format conversions of that work. Base model is Comfy-Org/MiniMax-H3.

Requires BF16

These are BF16 and expect a BF16 pipeline. They will not behave correctly in FP16.

Call it 6 steps, not 4

Upstream is named "4step". In practice 4 steps is a little weak — 6 steps is the sweet spot. The files here are renamed accordingly; the weights are unchanged.

Files

file	for base checkpoint
minimax_h3_turbo_6step_ema.safetensors	minimax_h3_fl2va_bf16
minimax_h3_turbo_6step.safetensors	minimax_h3_fl2va_bf16
minimax_h3_turbo_6step_ema_fl2va_pruned.safetensors	minimax_h3_fl2va_pruned_int8_convrot
minimax_h3_turbo_6step_fl2va_pruned.safetensors	minimax_h3_fl2va_pruned_int8_convrot

Start with the EMA one. It is noticeably cleaner than the non-EMA variant.

The _fl2va_pruned pair are remapped for the pruned int8/convrot checkpoint rather than the full BF16 one. Match the LoRA to the checkpoint you actually loaded.

Also works on the ref model

Trained for fl2va (first/last-frame), but it works on the ref2va (reference-to-video) path too, even though it was never trained for it. Pleasant surprise — worth trying.

Recommended settings

Sampler euler, scheduler beta57, 6 steps.

Strongly recommended: schedule the LoRA strength rather than pinning it flat. Using SCG LoRA Scheduler from SanDiegoDude/scg-utils:

setting	value
strength_start	4.00
strength_end	1.00
start_percent	0.100
end_percent	0.990
interpolation	linear
cutoff_outside_window	true
strength_clip	1.00

Ramping down from 4.0 to 1.0 across the run gives the best output I've found so far.

License

Apache-2.0, same as upstream.