MiniMax-H3-Image generates 1600万ピクセル images locally

MiniMax-H3-Image is a spin-off of the company's MiniMax H3 video model, and the pitch is aimed squarely at resolution. The usual route to a big image is to generate small and then hand it to an upscaler; MiniMax says its model produces 1600万ピクセル directly at a practical speed. That matters most for images where fine detail carries the meaning — text baked into a picture, textures added in editing, or interior mockups built from reference photos. The model accepts up to 16 reference images, and MiniMax showed examples including filling in terrain around a spot pinned on a map.

The distribution choices are as telling as the specs. The open weights are on Hugging Face under the MiniMax-H3-Image COMMUNITY LICENSE AGREEMENT, with free individual use in Japan. ComfyUI already supports it natively, with official workflows for both generation and editing — but those workflows default to 4 megapixels, so anyone wanting to judge the model's real capability has to switch the setting to 16 megapixels. For teams that run models on their own GPUs, that combination of open weights and existing tooling lowers the barrier to trying it.