The BEST local AI video generator is here! Minimax H3 tutorial

MHMiniMax H3MiniMaxUnverified

Open-source multimodal video generator with built-in audio that runs locally in ComfyUI. Supports text-to-video, image-to-video and reference-to-video (images, video, audio), with strong prompt following and character consistency. Roughly one minute of compute per second of video; pruned versions reportedly run on 12 GB VRAM. Licence is permissive for commercial use under $20M revenue, but excludes the EU, UK, Korea and US unless a separate licence is obtained.

COComfyUIComfy OrgAvailable in EU

Node-based platform used to run MiniMax H3 locally. The creator shows updating it, loading the MiniMax templates and wiring the text-to-video, image-to-video and reference-to-video workflows.

SASageAttentionAvailable in EU

Attention optimisation installed from prebuilt wheels matched to PyTorch, CUDA and Python versions. Used with the KJ Nodes patch node, it speeds up MiniMax H3 generation by about 20-30%.

COComfyUI-KJNodeskijaiAvailable in EU

Custom ComfyUI node pack that provides the Patch Sage Attention node. It is installed by cloning the repo and installing its dependencies.

COComfyUI-Spectrum-MiniMax-H3Available in EU

Custom ComfyUI node built for MiniMax H3 that speeds up generation further. It is installed by cloning the repo into custom_nodes, then adding the Spectrum Apply node before the guider and scheduler.

WAWan2GPAvailable in EU

Low-VRAM platform whose author says MiniMax H3 runs on as little as 5 GB VRAM at 480p. The creator points to a previous install tutorial.

ATAI ToolkitOstrisAvailable in EU

LoRA training toolkit that has already added support for MiniMax H3, so users can train their own LoRAs. The creator calls the topic technical and out of scope.