MiniMax-H3 Dual Turbo Lora
hybrid adapters for fast video generation
1 min readupdated
This repository contains two custom merged LoRA adapters for the MiniMaxAI/MiniMax-H3 architecture. It combines work from three communities for ultra-fast inference in just a few steps, temporal consistency, and simultaneous generation of synchronized audio and video. Built by linearly interpolating the weights of the base adapters, the pack is designed to work with diffusers and ComfyUI. Useful for content creators looking for ways to speed up generating video with sound.
Key features
- Ultra-fast inference in 5–9 sampling steps
- Synchronized audio and video generation
- Merges three cutting-edge LoRA checkpoints via weight interpolation
- Full compatibility with the Diffusers and ComfyUI ecosystems
What you can do
- Generate short video clips with synchronized audio in ComfyUI
- Speed up MiniMax-H3 generation using the recommended 6 sampler steps
- Experiment with two weight variants (.safetensors) to dial in the best quality
Example
Load the adapter on top of the standard MiniMaxH3Transformer3DModel base model in ComfyUI, set the sampling steps to 6, and set CFG to 1.0 using the lcm + simple scheduler.
Read more
DetailsBase model: MiniMaxAI/MiniMax-H3
Recommended steps: 6 (range 5–9)
CFG Scale: 1.0
Precision: bfloat16 / fp16
License: MiniMax H3 Community License Agreement (subject to geographic restrictions in the US, EU, UK, and South Korea)














Comments