MiniMax Music 3 and H3: ComfyUI music video tutorial

This tutorial connects MiniMax Music 3 to MiniMax H3 in ComfyUI to create a music video. The presenter starts with model downloads and a text-to-music workflow, then uses generated audio and a character reference image to build individual video scenes.

The speaker describes Music 3 as an open-weight model with a five-minute song limit and 32 kHz stereo output. The stated 8 GB VRAM minimum applies to the official weights; quantized ComfyUI versions are presented as options for lower-memory hardware. Setup covers the diffusion model, text encoder and VAE, with precision choices for the first two.

The M3 Songwriter node turns a song idea into a caption and tagged lyrics for review before audio generation. Controls include genre, vocals and duration. The presenter reports pronunciation problems in some languages and recommends allowing 80 or 90 seconds for lyrics planned around 60 seconds, since instrumental passages take time too.

For video planning, the Hermes AI agent uses DeepSeekV4 Flash for orchestration and an audio-understanding MCP. The H3 Video Prompt Enhancer supplies scene prompts. The demonstration then copies those prompts into ComfyUI, connects the reference audio input and aligns audio offsets with each scene. Six 15-second clips form the example. The presenter advises reviewing and regenerating clips when audio falls out of sync.