MiniMax H3 Turbo Ref2VA 8-Step v1.0: Reference-to-Video LoRA
Download minimax_h3_ref2v_turbo_8step_v1.0_comfyui_bf16.safetensors: Lightx2v's 8-step 768p Ref2VA distillation for MiniMax H3 with audio, plus updated example workflows.
The Lightx2v team and ModelTC updated Minimax-h3-Turbo with Ref2VA Turbo 8-step v1.0 768p, an 8-step 768p distillation of the reference-to-video path: one or more reference images drive a generated video with audio, at a fraction of the base model's step count (release repo).
This fills the last gap in the Turbo lineup. The FL2VA side already had 8-step and 768p checkpoints since the v1.0 release, while Ref2VA had only the 4-step v0.1 at 544p.
Model specs
The model spec table now lists six checkpoints across FL2VA and Ref2VA. The new entry:
| Model | Tasks | Training resolution | Training shifts (video / audio) | Distillation steps (NFE) | Recommended inference steps |
|---|---|---|---|---|---|
| Ref2VA Turbo 8-step v1.0 768p | Ref2VA | 768p (1344x768) | 6 / 3 | 8 | 8 |
Two files ship for it on Hugging Face:
minimax_h3_ref2v_turbo_8step_v1.0_768p_bf16.safetensorsfor Diffusersminimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensorsready for ComfyUI, no conversion needed
LightX2V Studio currently serves the FL2VA Turbo 8-step v1.0 768p LoRA; the same repo now covers the Ref2VA path at 8 steps and 768p.
ComfyUI usage
MiniMax H3 has native ComfyUI support, so no custom node is required for the LoRA:
- Download
minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensorsintoComfyUI/models/loras/. - Load the base MiniMax H3 checkpoint in a Ref2V workflow and apply the LoRA between the model loader and the sampler.
- Connect one or more reference images through the reference-input branch and run 8 sampling steps at 768p (for example 1344x768).
The official MiniMax H3 R2V workflow template is a good starting point. The project's own example graphs are also up to date: video_minimax_h3_ref2v_lightx2v_turbo.json defaults to the Ref2VA Turbo 4-step v0.1 checkpoint, and the T2VA/I2VA graphs default to the FL2VA Turbo 8-step v1.0.
For reference-image sizing, the project recommends the match mode used during distillation training, exposed through --reference-resize-mode in its inference entry point.
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 and the LightX2V API run the Turbo line in the cloud
Comments
Sign in with GitHub to join the discussion.