H3-Image
MiniMax-H3-Image is an open-weight image generation and editing model developed by MiniMax. Built upon a 33-billion-parameter diffusion transformer architecture, the model handles both text-to-image creation and multi-image editing within a single unified checkpoint. It is designed to generate native high-resolution images up to 16 megapixels directly at various aspect ratios without requiring a separate upscaler pass. A defining feature of the model is its inference efficiency. The checkpoint is distilled down to just 6 sampling steps without classifier-free guidance, enabling it to produce high-quality outputs in under 10 seconds. Furthermore, the architecture supports up to 16 reference images, allowing users to supply multiple style references, character sheets, and scene layouts simultaneously for granular editing and consistent composition tasks. The model excels in dense text rendering, subject identity preservation, and complex multi-reference workflows. By leveraging its extensive pre-training and post-training pipeline spanning data optimizations, reinforcement learning, and streamlined sampling, MiniMax-H3-Image bridges the gap between fast generation speeds and production-grade visual detail. It offers robust support for creative concept art, product visualization, and layout iterations while maintaining precise adherence to text prompts and visual guidance.
Create with Crafiq
Generate images, 3D models, video and audio in one studio.
Explore the studioHow H3-Image ranks
H3-Image is highlighted in the table below. Switch the metric to see how the ordering changes.