MiniMax H3 Turbo LoRA v1.0: 8-Step and 768p FL2V Weights
Lightx2v and ModelTC release MiniMax H3 Turbo LoRA v1.0 with an 8-step FL2V checkpoint and a 4-step 768p variant, both with ComfyUI-ready bf16 weights.
On August 11, the Lightx2v team and ModelTC released v1.0 of their MiniMax H3 Turbo LoRA, following the early v0.1 preview from August 7. The update adds two new FL2VA checkpoints in bf16, each shipped with a ComfyUI-ready format, including a variant trained at higher resolution.
Example frame from the Lightx2v Turbo LoRA project's demo output
What's in v1.0
| Model | Tasks | Training resolution | Distillation steps | Recommended inference |
|---|---|---|---|---|
| FL2VA Turbo 4-step v0.1 | FL2VA / T2VA | 544p (mixed AR) | 4 | 4 |
| FL2VA Turbo 8-step v1.0 | FL2VA / T2VA | 544p (mixed AR) | 8 | 8 or 4 |
| FL2VA Turbo 4-step v1.0 768p | FL2VA / T2VA | 768p (1344x768) | 4 | 4 |
The v1.0 release consists of two new checkpoints on Hugging Face:
- 8-step v1.0 (
minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors): distilled at 8 NFE with 12/3 video/audio shifts, and usable at 8 or 4 inference steps. - 4-step v1.0 768p (
minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors): a 4-step distillation trained at 768p (1344x768) with 6/3 shifts, aimed at higher-resolution first/last-frame-to-video generation.
Each checkpoint has a ComfyUI-format twin (..._comfyui_bf16.safetensors) hosted in the same repository, so no separate conversion is needed.
ComfyUI usage
MiniMax H3 has native ComfyUI support, so no custom node is required for the LoRA:
- Download the ComfyUI-format file (for example
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors) intoComfyUI/models/loras/. - Load the base MiniMax H3 checkpoint in a FL2VA workflow and apply the LoRA between the model loader and the sampler.
- For the 8-step checkpoint, run 8 sampling steps (4 also works); for the 768p variant, run 4 steps at 768p resolution.
The official MiniMax H3 image-to-video workflow template is a good starting point: insert the Turbo LoRA between the model loader and the sampler and switch to the step settings above.
Availability
- Weights: lightx2v/Minimax-h3-Turbo on Hugging Face (bf16 diffusers and ComfyUI formats)
- Code: ModelTC/Minimax-H3-Turbo on GitHub, with a batch inference script and NFE/LoRA comparisons
Comments
Sign in with GitHub to join the discussion.