MiniMax H3 and ComfyUI: local 3D rendering setup

Learn to turn 3D playblasts into AI video with MiniMax H3 in ComfyUI, including 24 FPS inputs, reference images and audio for lip sync.

Player not loading? Watch on YouTube

This tutorial explains a ComfyUI workflow that uses MiniMax H3 to turn a 3D viewport playblast into a rendered video. The presenter describes it as an open source option for local AI rendering and demonstrates generated sound alongside imported dialogue for lip sync.

Setup covers importing the workflow, installing missing custom nodes, restarting ComfyUI and placing model files in the folders indicated by the workflow. The presenter points to a pruned model download of around 12 GB for older GPUs; that figure describes the model file, not a stated VRAM requirement. An optional speedup group uses a LoRA and caching. He prefers eight steps with it, reports worse quality at four, and recommends bypassing the group and using 20 steps for maximum quality.

The workflow reformats input video to supported dimensions and clip lengths. The tutorial specifies 24 FPS and shows Blender workbench exports with clay shading or outlines. Prompts describe the reference video's role, with optional timelines and audio directions. Character and environment images provide additional visual guidance.

Examples reveal limits as well as useful results. Sparse backgrounds can cause sliding, and extreme camera moves can break scene consistency. For dialogue, the presenter says the imported audio must match the clip length. A Patreon version adds reference inputs and SeedVR 2 upscaling to 2K or 4K.