Summary
The Nerdy Rodent channel walks through running MiniMax H3, a newly released open-weight video generation model, locally on a home PC using ComfyUI. The tutorial covers four distinct generation modes: plain text-to-video, text plus a starting image, first-and-last-frame transformation, and a reference model that accepts multiple image and video inputs simultaneously — up to six combined references in the examples shown.
The presenter tests the FL2VA and reference model variants on an NVIDIA RTX 3090 and confirms that ComfyUI supports the model down to an RTX 3060 with sufficient system RAM. Output resolutions demonstrated range from 864×480 up to 1216×672, and the video notes that pushing beyond the documented 15-second maximum does not produce garbled output. One standout test involves feeding four images, a video clip, and extracted audio as simultaneous reference inputs to generate a scene with multiple characters — a capability that sets H3 apart from earlier locally-runnable video models.
The workflow follows the channel’s signature “rodent method” of color-coded, modular ComfyUI node groups designed for readability and easy updates. Viewers are shown how to handle prompting conventions for multi-image references (using triangular bracket indexing), when to use the FL2VA versus the reference model, and practical tips on generation time expectations for consumer hardware.
📺 Source: Nerdy Rodent · Published August 04, 2026
🏷️ Format: Tutorial Demo
