MiniMax H3 Turbo LoRAs

MiniMax-H3 Turbo LoRA is a distillation-style LoRA adapter for the MiniMax-H3 joint video and synchronized stereo audio generation model. It lets the base model render in as few as 4 sampling steps instead of the usual ~20, roughly a 5x sampling speedup. The Hugging Face repository is published by larryvrh; YouTube coverage also mentions similar Turbo LoRAs from LightX2V, Joy Fox and Kijai, including a resized ~315 MB variant.

It is used in ComfyUI through the ComfyUI-MiniMax-H3-Turbo custom nodes (a Turbo LoRA node and a Turbo Sampler), inserted into the official MiniMax-H3 text-to-video or image-to-video workflow. The model card recommends the v4 step-600 EMA checkpoint for most work, 4–8 steps (6–8 for best quality), strength 1.0 and the simple scheduler. One LoRA file works with the full and pruned/quantized MiniMax-H3 bases; a low_vram option merges the weights to reduce peak memory.

Known limits per the model card: it is still a preview with training ongoing. At 4 steps with large, fast motion, v4 can show motion smear or ghosting, and the older v1 checkpoint may be preferable there. Going beyond 8 steps can add over-sharp artifacts, and audio quality is still being improved. The weights are bf16, around 744 MB per file, and you also need the base MiniMax-H3 model, VAEs and text encoder.