MiniMax H3 in ComfyUI: multi-shot setup and comparison

This tutorial explains MiniMax H3 Motion Director, a ComfyUI node for generating separately prompted shots in sequence. The speaker says it passes motion and audio context between shots while generating each segment independently. The claimed "infinite length" means adding more segments, rather than holding an entire long video in GPU memory. The demonstration uses a 5090 and 64 GB of system RAM; it does not establish a minimum hardware requirement.

Setup covers installing the node through ComfyUI Manager, restarting ComfyUI and selecting a compatible model. The workflow also needs the Qwen 3 VL 32-billion-parameter MiniMax text encoder, video VAE and audio VAE. The director mode must match the selected model. A four-step turbo LoRA reduces sampling from 25 steps to four, but the speaker reports visible quality loss, especially with fast motion. Chunking and attention settings address memory use and speed.

The local AI workflow includes shot durations, source images and selective regeneration. The speaker recommends keeping 24 FPS because changing it affects duration in this workflow. Comparisons use image-to-video with an initial image and reference-to-video with an image or character sheet in every segment. The speaker finds better character consistency with repeated references, while the image-to-video example shows facial drift after a turn.