MiniMax-H3 Turbo LoRA ⚡
Few-step MiniMax-H3 Turbo LoRAs for video and synchronized audio — 4–8 steps instead of dozens — served through Hugging Face Inference Providers. Sign in so generation is billed to your own account.
Text-to-video
Ref-to-video
Audio-video
T2VA FL2VA Ref2VA
Try it
Generation runs on a third-party inference provider (WaveSpeed). A few seconds of video costs a few cents; you can watch usage on your HF account after the first request.
Example outputs
Run it in ComfyUI
This Space demos the LoRAs over an inference provider. To run them locally with ComfyUI, grab the recommended v4 step-600 EMA pruned LoRA and the matching workflow:
- Base model: Comfy-Org/MiniMax-H3 (pruned/curve-form variant)
- LoRA: minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors →
ComfyUI/models/loras/ - All LoRA variants: model files
| Setting | Value |
|---|---|
| Steps | 6–8 |
| Sampler / Scheduler | Euler / Beta |
| LoRA strength | 1.0 (adjust 0.8–1.2) |
| Video sigma shift | 12 |
| Audio sigma shift | 4–6 |