MiniMax H3-Image - Bringing 4K image generation into the 10-second era
MiniMax H3-Image is an open-weight model for native 16MP image generation and editing. It uses just six diffusion steps, handles text-to-image and editing in one checkpoint, and supports up to 16 reference images.
Replies
Hi everyone!
@MiniMax is bringing its H3 architecture to image generation!
16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds.
The weights are already up on Hugging Face, and you can also use it through the API or Hailuo AI.
For local deployment, the image checkpoint alone is about 69GB, and it uses a separate H3 text encoder of roughly 67GB.
The 16 reference images is the part that caught my eye — keeping a character or product consistent across edits is the thing I struggle with most when I generate visuals. 6 diffusion steps for native 16MP in 10 seconds is wild too. The 69GB local checkpoint is a bit much for my machine though, so I'd probably start with the API and see how well it holds consistency across a set of images.
16 reference images in one open weight checkpoint is wild. congrats minimax team