MiniMax H3: ComfyUI setup and video workflows

This tutorial covers MiniMax H3 setup in ComfyUI and its text-to-video, image-to-video and reference-to-video workflows. The presenter uses Windows portable ComfyUI and distinguishes local templates from API-tagged templates that run in the cloud.

Setup requires a diffusion model, the Qwen-3-VL-32B text encoder, and separate audio and video VAEs. The listed diffusion downloads range from 66 GB for the full model to 34 GB for INT8 and 21 GB for pruned variants. The presenter notes a quality tradeoff with compression and says the text encoder can use system RAM. Reports of operation with 12 GB VRAM come from other users, rather than a test shown here.

Examples cover generated audio, multiple image references, video background replacement and a song used as an audio reference. The presenter reports roughly one minute of processing per second of output on their setup. Optimization instructions cover SageAttention, KJ nodes, Easy Cache and ComfyUI Spectrum. The demonstrated environment uses PyTorch 2.6, CUDA 12.4 and Python 3.11; SageAttention wheels must match those versions.

For lower-memory local AI use, the presenter cites Wan2GP's author as claiming 480p generation with 5 GB VRAM. The closing license discussion describes a $20 million annual commercial revenue threshold and separate licensing for the EU, UK, Korea and US. These are the speaker's account of the terms, not an independent license review.