MiniMax H3 in ComfyUI: setup and batch video tutorial

This tutorial explains a ComfyUI workflow the speaker calls MiniMax H3, which reads scene prompts from a folder, matches their reference IDs, queues clips and merges the results. The speaker calls the opening minute zero-shot with ChatGPT-written prompts, though later also describes prompt refinement and manual improvements. Although the title advertises two-hour movies, the walkthrough focuses on this shorter example.

Setup covers Windows installation, custom-node bundle 100 and model downloads through the SwarmUI model downloader. The speaker recommends Python 3.12 and INT8 models, describing INT4 as an option for lower VRAM. They say the model can run on a 6 GB GPU with slower generation. The demonstration uses cloud GPUs on Massed Compute; RunPod and SimplePod installers are briefly mentioned. Suitable local hardware is an alternative, and prompts can be written manually instead of using ChatGPT.

For local AI video projects, the suggested process is to render a 0.4 MP draft, review the sequence and refine prompts before the final render. The speaker recommends 1344x768 for this workflow and describes a 30-second clip limit, with 15 seconds preferred and 10-second scenes used here.

The speaker gives 99 loaded references as an example, while saying each generation accepts up to three audio files, three video clips and nine images. Extra audio or video references are discarded. The speaker reports weaker results with audio references than images and explains how to map image IDs to characters. A weak scene can be regenerated separately and replaced.