MiniMax releases H3-Image weights for 4K generation and reference editing

MiniMax announced open weights for H3-Image on November 11 and made the image model available through its API. H3-Image generates still images from text and edits them with references. A game team can use a character sheet or level map to explore what a scene might look like before building it in an engine. MiniMax announcement

Give each reference a role

The image-editing endpoint takes a prompt and one to 16 reference images. You could use one image for the character’s outfit, another for the palette and a third for the scene layout, then state which details each reference should guide. The API caps the entire request body at 64 MiB and accepts references to previously uploaded files. Image-editing documentation

In MiniMax’s model card, an illustrated map with a marked camera direction becomes a first-person still looking toward a bridge and waterfall. A level team could use that kind of image to discuss framing and landmark visibility before blocking out the route in 3D. MiniMax showcase

Check the API bill and the local download

The hosted API’s 4K area setting targets 4096 × 4096 pixels for a square image. Other aspect ratios have different dimensions. Standard and Highspeed have separate prices for text input, image input and image output. MiniMax lists a launch discount of 50% for API requests with delivered images of at least 16.6 million pixels until November 17 at 16:00 UTC. API guide · Pricing

For local use, the model card lists a 68.65 GB H3-Image package and a 66.7 GB H3 text encoder that downloads separately if it is missing from the cache. The same card links deployment recipes and the model’s community license. Model card · License

See 2D Art & Animation and World & Level Building for related production workflows.