MiniMax H3 Local Install — Text-to-Video and Multimodal Reference Workflow (Patreon)
Downloads
Content
Video : https://youtu.be/RIFTcIZqaS4
Hey everyone.
New video is up: MiniMax H3 Might Be The Best Local AI Video Now! Multimodal Reference2Video Walkthrough.
MiniMax finally dropped H3 open weights on Hugging Face and ModelScope, and ComfyUI already has merged / quantized files ready for consumer GPUs. This is not just another text-to-video model. It is an omni-modal setup — text, images, video, and audio — and in the video I walk through what actually runs locally right now (H3 base / 768p path), plus the honest limits vs Seedance-style fight-scene character replace.
Free workflows attached for the three modes I demoed:
1. video_minimax_h3_t2v.json — text-to-video (FL2VA)
2. video_minimax_h3_i2v.json — image-to-video / first-frame style (FL2VA)
3. video_minimax_h3_r2v.json — multimodal reference-to-video (Ref2VA)
These use the pruned Int8 Convrot weights + Qwen3-VL 32B NVFP4 text encoder, with separate video and audio VAEs. Good starting point if you want to try H3 without hunting every file yourself.
Paid post has the fuller breakdown — what each graph is doing, model file list, frame-length formula notes, tips, and common issues from the recording session.
— Benji
MODELS REFERENCED
Video : https://youtu.be/RIFTcIZqaS4
- MiniMax H3 (official weights): https://huggingface.co/MiniMaxAI/MiniMax-H3
- ComfyUI merged / quantized files: https://huggingface.co/Comfy-Org/MiniMax-H3
- Diffusion variants: FL2VA (text-to-video / image-to-video), Ref2VA (reference-to-video + audio)
- Text encoder: Qwen3-VL 32B (BF16 / Int8 Convrot / NVFP4)
- VAEs: MiniMax H3 video VAE + audio VAE (separate files)
- Note: Distillation LoRA not available yet at time of recording; default sampling ~20 steps