ComfyUI MiniMax H3 setup and video workflow tutorial

This tutorial covers MiniMax H3 video generation in ComfyUI, with workflows for text prompts, first and last frames, multiple reference images, and speech or singing from an image plus audio. Setup includes updating ComfyUI and Pixaroma Nodes, placing model files in their folders, and loading a 32-billion-parameter Qwen text encoder and two VAEs.

The presenter recommends pruned INT8 versions of the separate first/last-frame and reference models. The Sage Attention installation shown targets Windows with NVIDIA GPUs; he reports problems with 20-series cards and expects 30-, 40-, and 50-series cards to work. He also fixes a failed run by removing a startup argument that disabled Dynamic VRAM.

Workflow controls cover aspect ratio, dimensions rounded to multiples of 32, and duration up to 15 seconds. Custom ChatGPT prompt templates help describe motion and reference inputs. The presenter compares generation times across resolutions, but these are results from his setup rather than general performance guarantees.

Limitations receive attention too. He could not get video-reference inputs working properly, noticed occasional lines in audio-driven output, and found that image edits could alter unintended details. His license discussion identifies excluded territories and describes his application approval; his comments about personal use are his interpretation.