MiniMax open-sourced the H3-Image generation model, producing 4K images in 10 seconds
MiniMax announced the release and open-sourcing of H3-Image, derived from H3 with text-to-image, image reference and image editing capabilities, generating 4096 by 4096 pixel 16MP images within 10 seconds, supporting up to 16 reference images with a half-price launch promotion for 16MP and above.
MiniMax announced late on November 10 that H3-Image had been released with open weights. It is an image generation model derived from H3 with text-to-image, image reference and image editing capabilities. The company says it can generate a 4096 by 4096 pixel image of about 16MP within 10 seconds, while presenting rich realistic detail and offering strong instruction following and multi-reference image understanding. MiniMax said that building on the broad image capabilities granted by H3 pre-training, H3-Image further strengthened the entire post-training pipeline, with work on data, reinforcement learning, step compression, positional encoding and deployment optimization, giving the model competitiveness in speed, image quality and instruction following at once. Compared with the mainstream 1 million or 4 million pixel outputs, 16 megapixel images offer markedly better detail, and H3-Image produces 16MP output in 10 seconds at up to 5 times the speed of mainstream offerings. The model provides strong instruction-following ability among open models and supports up to 16 reference images. In its first week of release, MiniMax offers a half-price promotion on the official price for generation requests at 16MP and above. Weights are open on Hugging Face under the MiniMaxAI organization, with official API and MiniMax Code access also provided.