MiniMax H3 in ComfyUI: four video workflows

This tutorial demonstrates a MiniMax H3 all-in-one workflow for ComfyUI. It covers text-to-video, image-to-video, reference-based generation and lip sync, with shared multimedia inputs, acceleration controls and a live preview.

The text-to-video example uses a 10-second clip at 1280 by 736 pixels. The presenter reduces generation from 20 steps to eight and checks a preliminary preview before letting the workflow finish. An RTX upscale node then processes the result to 2K. The presenter's praise for motion and prompt adherence reflects these examples rather than a controlled benchmark.

For image-to-video, one image supplies the starting frame; a second supplies the final frame. Although the loader accepts more images, this path uses only the first two. The reference-to-video example instead uses eight images for a 15-second scene, including a storyboard made with ChatGPT. Audio and video references are supported but bypassed in that test.

The lip-sync example combines three singer images, a song and a reference video intended to guide camera movement. A dedicated prompt module guides the performance. Viewers can follow the workflow controls, but the supplied material gives no hardware requirements or measured generation times for planning a local AI setup.