MiniMax H3 Turbo LoRA: 4-Step FL2V Distillation for ComfyUI
Lightx2v and ModelTC distill MiniMax H3 into a 4-step FL2V Turbo LoRA, cutting sampling from ~20 steps to 4 with a ComfyUI conversion and demo Spaces.
The Lightx2v team and ModelTC released a 4-step Turbo LoRA for MiniMax H3, distilling the video model's first-last-frame-to-video (FL2V) generation from roughly 20 sampling steps down to 4 (Hugging Face, GitHub). The LoRA, about 1.3 GB, is an early preview: v0.1 focuses on FL2V quality, with a Ref2V distillation already on the roadmap.
A frame from the author's example output at 4 steps
What's in the release
- Weights:
minimax_h3_fl2v_turbo_4step_v0.1.safetensorson Hugging Face, distilled from the MiniMax-H3 base model - Inference code: the ModelTC repository provides batch MiniMax-H3 inference plus NFE/LoRA comparisons against the base model at 50 NFE
- Demo Spaces: akhaliq's MiniMax-H3-Turbo-Lora Space and an official-org Space let you try the LoRA without local hardware
- Roadmap: improve FL2V visual details and overall quality, then develop a Ref2V-based distillation
The release notes label v0.1 a preview: visual details still need improvement, and an updated version is in development.
ComfyUI usage
ComfyUI does not need a custom node for this LoRA. Kijai published a ComfyUI-compatible conversion in Kijai/MiniMax-H3_comfy (minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors, plus a resized bf16 variant) within an hour of the release. Drop the file into models/loras/, load the base H3 checkpoint, and apply the LoRA at 0.75 strength with 4 steps using an er_sde or sa_solver sampler, per Kijai's test notes.
The official MiniMax H3 text-to-video workflow template is a good starting point; insert the Turbo LoRA between the model loader and the sampler and switch to the 4-step settings above.
Note: the authors observed some differences between Diffusers and ComfyUI inference results (audio sounded more natural in Diffusers). An audio bug in the ComfyUI path was fixed in a nightly update. For reproduction matching the paper results, follow the ModelTC inference script.
Example output shared by the authors in the release discussion
Availability
- Weights: lightx2v/Minimax-h3-Turbo on Hugging Face
- Code: ModelTC/Minimax-H3-Turbo on GitHub
- ComfyUI conversion: Kijai/MiniMax-H3_comfy
- Demo: MiniMax-H3-Turbo-Lora Space
Comments
Sign in with GitHub to join the discussion.