MiniMax-H3-Image: Open-source DiT that generates and edits 16MP images in 6 steps

This model integrates image generation from text prompts and editing of existing images into a single checkpoint. It is based on the MiniMax H3 family of DiT architectures and features a pure image generation structure with the audio layer removed. It utilizes Qwen3-VL as the text encoder to materialize creative intent into visual worlds.

1 / 8

It directly generates images at up to 16-megapixel resolution without an upscaler. It supports various aspect ratios ranging from 4096x4096 to 6272x2688 at a 21:9 ratio, with inference completed in just 6 sampling steps. It maintains a high level of detail and realism even without classifier-free guidance, efficiently managing computational costs compared to existing methods.

It can accept up to 16 reference images for use in editing tasks. It supports major inference frameworks such as SGLang, vLLM-Omni, Diffusers, and ComfyUI, and runs on both NVIDIA and AMD accelerators. Local deployment requires 68.65GB of BF16 weights, and a prompt expansion model can be used optionally.

It is evaluated as having the strongest instruction-following capability among open-source models. Text generation, image editing, and prompt expansion can be called directly via API, and web app and desktop clients are also provided. It is suitable for creators and developers who require precise control and fast generation speeds.