MiniMax opens H3-Image weights for image generation and editing at up to 16 megapixels
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Detailed high-resolution image generation and editing can now run on local hardware without dependence on a closed API.
H3-Image combines text-to-image generation and editing from one or more references in a single checkpoint. The model directly generates images of about 16 MP without an upscaler, including 4096×4096 and aspect ratios up to 21:9, and accepts up to 16 references. The distilled version uses six sampling steps; its diffusion transformer has 33.1 billion parameters. Weights and configurations for SGLang, vLLM-Omni, diffusers and ComfyUI are published on Hugging Face, with support for Nvidia and AMD accelerators. MiniMax claims 4K generation in under 10 seconds; speed depends on the hardware configuration. Local use also requires the MiniMax-H3 text encoder, and the company offers a separate model for expanding short prompts.
Sources 2