MiniMax H3 Runs Locally in ComfyUI With Native Stereo Sound
A community ComfyUI workflow packages MiniMax H3 video generation with a Turbo LoRA and resolution controls, letting modest local GPUs run the open-weight model.
Takeaways
- ComfyUI workflow for MiniMax H3 video model, 619 stars, tuned for low-VRAM consumer GPUs.
- Single JSON supports first-last frame, image-to-video, and text-to-video via input bypassing.
- Bundles Turbo LoRA for 4-step distilled sampling, defaults to 8 steps.
- Uses pruned INT8 base checkpoint and NVFP4 AWQ-quantized Qwen3-VL 32B text encoder.
- Control panel exposes aspect ratio, megapixels, duration, steps, and seed on outer graph.
- Outputs 24 fps video with synced audio, frame count aligned to 17k+5.
The ComfyUI workflow packages MiniMax H3 text-to-video, image-to-video, and first-last-frame generation into one graph. Its exposed controls handle model loading, resolution, duration, sampling, and synchronized audio decoding, giving developers a practical route to local experimentation with quantized weights.
One graph, three entry points
MiniMax H3 is a multimodal video model that can process text, images, video, and audio as shared context. According to the H3 model page, it generates video with native stereo sound, supports output up to 2K, and accepts integer durations from 4 to 15 seconds. MiniMax released the weights on August 2, 2026, enabling local pipelines such as this one.
The workflow groups model loading, Turbo LoRA application, sampling, resolution calculations, and audio-video decoding inside a subgraph. A single file, video_minimax_h3_t2v_0909.json, exposes three generation modes:
| Mode | Inputs | Behavior |
|---|---|---|
| First-last-frame video | first_frame and last_frame |
Generates the motion and audio between two supplied images. |
| Image-to-video | first_frame |
Animates one source image while bypassing the final-frame input. |
| Text-to-video | Prompt only | Bypasses both image inputs and generates from text conditioning. |
The supplied JSON exposes text and still-image conditioning. H3’s broader multimodal input capabilities require other graphs or nodes.
Why a four-step LoRA runs at eight
The main speed optimization comes from the Turbo node pack. It adds a MiniMaxH3TurboLoRA node and loads minimax_h3_turbo_4step_ema_ckpt850.safetensors.
A LoRA is a compact adapter that changes how a larger model behaves without replacing its base weights. This adapter was distilled for four-step sampling, reducing the number of denoising passes needed to form a clip. The bundled workflow uses eight steps with the