MiniMax H3 in ComfyUI: latent upscale tutorial

This tutorial covers a two-stage MiniMax H3 video workflow in ComfyUI: generate at a lower resolution, then use an additional latent upscale model for the second sampling pass. The presenter typically starts at 0.6 or 0.7 megapixels and sets the demonstrated upscale workflow to a maximum of 2 megapixels. This is a community implementation, not MiniMax's official 2K upsampler.

The presenter explains changes in their fork of ComfyUI-MMH3-UltimateUpscale, including an option to keep the roughly 600 MB upscaler resident in memory. They report unnecessary loading and offloading on GPUs with 24 GB or more VRAM, while describing chunk offloading as useful for cards with 12 GB or less. These are observations from their setup, rather than guaranteed hardware requirements.

The walkthrough covers precision choices, optional temporal and spatial splitting, and matching the second-stage width and height connections. The specialized sampler needs the latent upscale model input. Text-to-video and reference-to-video examples show how the conditioning carries into the second pass for local AI video generation.

Results vary: the presenter calls one fight scene a failure and finds another attempt more acceptable. The closing setup advice places the model in the latent upscale models folder under models, rather than the upscaler or diffusion model folders. The demonstrated file uses BF16; FP32 is suggested when memory permits.