MiniMax H3 in ComfyUI: local video setup and workflows

This tutorial covers MiniMax H3 setup in ComfyUI and its text-to-video, image-to-video and reference-to-video workflows. The presenter describes H3 as an open source model that generates clips up to 15 seconds long at 24 fps with native stereo audio. He distinguishes local workflows from paid API templates, which he says are required for 2K output at the time of recording.

Model selection depends on GPU memory. The presenter uses INT8 on a 32 GB card and expects it to work above 24 GB. He recommends pruned INT8 for an RTX 3060 with 12 GB using offloading, while suggesting more than 32 GB of VRAM and sufficient system RAM for BF16 pruned. He describes FP8 as tuned for faster operation on RTX 40 and 50 series cards. These are his hardware recommendations, rather than measured minimum requirements.

The workflow walkthrough explains aspect ratio, resolution settings and timestamped prompts for multiple shots, dialogue and audio cues. Image examples include assembling a poster and animating a web page. Reference-to-video accepts images, video and audio, with a stated limit of 12 media inputs; prompts should refer to them in connection order.

The presenter also reports a consistency failure: a fantasy creature becomes a regular tiger despite instructions to preserve its appearance. GGUF versions and further optimizations are discussed as future developments.