MiniMax H3 Turbo LoRA v4: 2-Step Audio-Video Generation
The v4 step-600 Turbo LoRA distills MiniMax H3 audio-video generation to 2-3 sampling steps, with updated ComfyUI nodes supporting pruned checkpoints and a bundled workflow.
larryvrh released v4 of the MiniMax H3 Turbo LoRA, and the few-step race keeps accelerating: the new minimax_h3_turbo_v4_step600_ema.safetensors checkpoint renders natively at 3 steps, and community tests pushed a working 2-step pass with clean audio (Hugging Face). It replaces the earlier ckpt500 / ckpt850 files as the recommended default, with markedly better micro-detail and the over-sharpened "plastic" look of the earlier v1 line fully resolved.
Example output shared in the Turbo LoRA conversion repository (drbaph)
What's new in v4
The v4 release is a new training recipe rather than just another step count. At 600 training steps it is the strongest checkpoint larryvrh has published so far:
- Static-frame enhancement — a big win for static and small-motion content
- Better micro-detail — faces, fingers and fine texture read more cleanly
- No more over-sharpening — the earlier v1 (~850) line's plastic look is gone
The one trade-off shows up only at 4 steps with large, fast motion, where v4 can produce motion-smear / trailing ghosting. Using 6-8 steps largely removes it; for the specific case of 4 steps plus heavy motion, the older v1 (~850) checkpoint remains the friendlier pick. The v4 line also tolerates higher step counts better than v1, which tends to over-sharpen at high steps with strength 1.0.
Steps and settings
The README recommends 4-8 steps (4 is the minimum, 6-8 look noticeably better), strength 1.0, and the simple scheduler. Community testing in the MiniMax H3 Discord landed on a similar sweet spot at 0.6-0.8 strength with 6-10 steps, and confirmed that the 2-step pass works with audio intact (video output is a bit softer). Training continues, with audio and fast-motion behaviour flagged as the two areas still being improved.
ComfyUI usage
The updated ComfyUI-MiniMax-H3-Turbo custom nodes now cover every base model: one LoRA file works with the full bf16 / int8 checkpoints and the pruned curve-form variants — the node auto-detects a pruned base and re-injects the time-conditioning at run time. A low_vram switch merges the LoRA into the weights for the lowest peak VRAM, and the bundled MiniMax-H3 Turbo Sampler auto-adapts to your ComfyUI version (recent builds handle the dual video/audio flow schedules natively via ModelSamplingAV, older builds use the node's own path).
A ready-made text-to-video workflow ships in the node repo:
Pruned checkpoints (drbaph)
For ComfyUI's built-in MiniMax-H3 LoRA loader, drbaph published pruned-model conversions of the v4 weights at MiniMax-H3-Turbo-Lora-ComfyUI: minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors (recommended) plus the non-EMA variant, with an example workflow (fl_minimax_h3_turbo_lora_example_workflow.json). His recommended starting configuration is 6-8 steps, Euler sampler, Beta scheduler, strength 1.0.
Demo output from the recommended v4 step-600 EMA workflow, shared by drbaph
Availability
- Weights: larryvrh/MiniMax-H3-Turbo-Lora —
minimax_h3_turbo_v4_step600_ema.safetensors(recommended),minimax_h3_turbo_v4_step600.safetensors, plus earlier v4 / v1 checkpoints - Custom nodes: Larryvrh/ComfyUI-MiniMax-H3-Turbo (installable via ComfyUI-Manager, search "MiniMax-H3 Turbo")
- Pruned conversions: drbaph/MiniMax-H3-Turbo-Lora-ComfyUI
Comments
Sign in with GitHub to join the discussion.