Video and a synchronized stereo soundtrack out of a single denoising pass, in as few as four sampling steps instead of twenty. The model card carries real guidance about which checkpoint to use and when — this turns that prose into a decision you can make in two clicks, and hands you the command.
minimax_h3_turbo_v4_step600_ema.safetensors
v4 · step 600 · EMA — the current best checkpoint.
| Setting | Value | What happens if you leave the range |
|---|---|---|
| Steps | 4–8, 6–8 preferred |
Below 4, the sampler has too little to work with. Past 8 it stops helping and starts introducing over-sharp artifacts. |
| LoRA strength | 1.0 |
Tuned for 1.0 across the whole 4–8 range. Only move it for a specific misbehaving clip: blurry ghosting → 1.05–1.2; over-sharp grain → 0.8–0.95. |
| Scheduler | simple |
Leave it. The custom sampler is built around it. |
| Base model | any |
Full (bf16, int8_convrot) and pruned variants all work — one LoRA file covers every base. |
generate.py is self-contained but still
needs a ComfyUI checkout for the H3 model, VAE and text-encoder definitions, plus the base weights from
Comfy-Org/MiniMax-H3.
.safetensors into
ComfyUI/models/loras/.SamplerCustomAdvanced from
MiniMax-H3 Turbo Sampler, scheduler on
simple, steps ≥ 4.This Space hosts nothing. Generation runs on a hosted provider through Hugging Face Inference Providers, which serve this exact LoRA. The provider is chosen automatically, honouring your own preference order. Sign in and the call is billed to your account — not to this Space.