10-Second 4K Image Generation: MiniMax Once Again Delivers Open-Source SOTA
MiniMax has released MiniMax H3 Image, an open-source image generation model that has quickly claimed state-of-the-art status among open-weight models. It natively outputs 4K images (around 16 million pixels) in roughly 10 seconds with the fast version, supports up to 16 reference images, and ranks first among open-weight models on Artificial Analysis for both text-to-image and image editing. The model is available via web, API, and local deployment on GitHub and HuggingFace.
In hands-on tests, H3 Image produced highly realistic WeChat chat screenshots with accurate iPhone status bar details, rendered complex graffiti-style text ("MiniMax H3 Image") with 3D lettering and spray-paint effects, and generated convincing Victoria's Secret backstage photos, cinematic street portraits, and Y2K-style retro portraits. It also handled image-to-image tasks well: turning a photo into a hand-drawn sticker, adding CCTV-style face bounding boxes and zoom-ins, and converting realistic images into anime collage styles. The model further demonstrated understanding of physical structure by generating a labeled exploded-view diagram of a 35mm camera, producing accurate outfit try-on images, e-commerce product posters, and typography-heavy Eastern botanical posters, and following complex multi-reference editing instructions.
Architecturally, H3 Image derives from MiniMax's full-modal video model H3, retaining its backbone and text encoder (a 32-billion-parameter multimodal language model) while dropping the audio branch. Its core is a 33.1-billion-parameter diffusion Transformer using 3D multimodal rotary position encoding, with text and reference-image tokens processed in a single stream and decoded by a VAE. Native generation avoids external super-resolution. The video-first lineage is credited with giving the model strong visual world knowledge, including viewpoint, material, lighting, and motion understanding.
The release has energized the global AI community: derivatives such as Pruna's P-H3-Image, Nvidia's Sol-H3, and Pika H3 Flex have appeared, with roughly 220 H3-based entries on Hugging Face. Chinese open-source technology is increasingly the foundation for overseas teams, and MiniMax is building a commercial loop around free weights plus paid hosted APIs and subscriptions. The article argues competition is shifting from raw model capability to ecosystem strength.
marsbit1h ago