MiniMax H3 FL2V Turbo 4-Step v1.2 LoRA: Cleaner Audio
Download minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors: Lightx2v's 4-step MiniMax H3 Turbo update with cleaner, more stable audio generation.
The LightX2V team released FL2V Turbo 4-step v1.2 768p, the newest entry in the Minimax-h3-Turbo distillation line. v1.2 keeps the same fast 4-step generation setup as v1.1 and focuses entirely on one weakness: audio quality. The release comparison shows cleaner and more stable generated audio with fewer artifacts at identical settings (release discussion).
What changed in v1.2
v1.2 is an audio-quality refresh of the FL2V (first/last-frame) 4-step 768p LoRA rather than a new architecture:
- Cleaner, more stable audio with fewer artifacts
- Same 4-step (NFE 4) generation setup as v1.1
- Both Diffusers and ComfyUI bf16 weights published side by side
| File | Format |
|---|---|
minimax_h3_fl2v_turbo_4step_v1.2_768p_bf16.safetensors | Diffusers |
minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors | ComfyUI, no conversion needed |
v1.1 vs v1.2 comparison (enable audio)
The team shared back-to-back comparisons using the same input, prompt, seed, and inference settings:
| v1.1 | v1.2 |
|---|---|
Recommended settings
| Setting | Value |
|---|---|
| Steps | 4 |
| Video shift | 6 |
| Audio shift | 3 |
| Sampler | Euler |
| Resolution | Up to 768p |
Video and audio shifts are set through the ModelSamplingMiniMaxH3 node, connected after the diffusion model loader. Note that some early testers on the Hugging Face discussion thread felt the visual quality took a hit relative to v1.1, so if video quality matters more than audio in your pipeline, keeping v1.1 is a reasonable choice.
ComfyUI usage
MiniMax H3 has native ComfyUI support, so no custom node is required for the LoRA:
- Download
minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensorsintoComfyUI/models/loras/. - Load the base MiniMax H3 checkpoint and apply the LoRA between the model loader and the sampler.
- Set 4 sampling steps, Euler sampler, video shift 6 and audio shift 3 via
ModelSamplingMiniMaxH3, then generate at up to 768p.
The project's own example graphs in ModelTC/Minimax-H3-Turbo are also up to date; the T2VA/I2VA graphs are the closest match for the FL2V path.
Availability
- Weights: lightx2v/Minimax-h3-Turbo on Hugging Face (Apache-2.0), bf16 Diffusers and ComfyUI formats
- Code and workflows: ModelTC/Minimax-H3-Turbo on GitHub
- Hosted: LightX2V Studio runs the Turbo line in the cloud
Comments
Sign in with GitHub to join the discussion.