What it does
MiniMax H3 writes the picture and the sound at the same time, from the same prompt, on your own machine. No separate music step. No voice tool. No editing the two together afterwards. You type a scene, you get a clip that already has its own audio.
What you need
Four files, about 42 GB in total, all free from the official Comfy-Org MiniMax-H3 repository on Hugging Face. Every filename below is a direct download link.
- minimax_h3_fl2va_pruned_int8_convrot.safetensors 21 GB, into
ComfyUI/models/diffusion_models/ - qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 15.7 GB, into
ComfyUI/models/text_encoders/ - minimax_h3_video_vae_fp16.safetensors 5.2 GB, into
ComfyUI/models/vae/ - minimax_h3_audio_vae_fp32.safetensors 0.6 GB, into
ComfyUI/models/vae/
The folder is the part people get wrong. A file in the wrong folder looks exactly like a missing model, and ComfyUI will not tell you which one it is.
No custom node packs. Everything in this file ships with ComfyUI.
You also need ComfyUI 0.30 or newer, and a card with 24 GB or more. Under that it will be slow and it may not run at all. Better to know now than after 42 GB of downloading.
How to use it
- Put the four files in their folders. Restart ComfyUI.
- Drag
stableyogi-minimax-h3-text-to-video.jsononto the canvas. - Write your scene. Describe the picture and the sound, so rain on a roof, a car going past, a voice. It listens to sound words.
- Press Run.
The clip lands in ComfyUI/output/video/.
The thing that catches everybody
Frame count is not free-form. H3 takes 17 times something, plus five:
- 73 frames is about 3 seconds
- 124 frames is about 5 seconds
- 192 frames is about 8 seconds
- 277 frames is about 11.5 seconds
Three more H3 workflows, all free
- Picture to video, start from a still you already made
- Turbo, the same clips in 8 steps instead of 20. One extra 1.9 GB file, nothing else changes
- Reference to video, give it pictures of a person and describe a scene they are not in
Written guides, free to read, no account needed: text to video and picture to video, explained, and turbo and reference to video, with the measured numbers.
There is a Pro version
Not another download. One file carrying every H3 mode on switches, with the speed settings already set the way the testing said to set them, including one that is switched off on purpose because turning it on cost fifteen seconds a clip here.
It comes with Pro membership, and you get it by finishing the MiniMax H3 course, six lessons, every workflow in this set explained node by node. The file unlocks on the last one. What Pro includes.