MiniMax drops open-weight 4K H3-Image

MiniMax tackles diffusion latency and editing fragmentation in one stroke by delivering 16-megapixel outputs in six sampling steps without secondary upscalers. Packaging high-resolution generation and reference-conditioned editing into an open checkpoint puts serious pressure on closed alternatives.

• –Distilled six-step sampling generates 4096x4096 images in under ten seconds, removing the latency bottleneck of traditional two-stage super-resolution pipelines.
• –Unified checkpoint handles text-to-image and image-to-image editing without swapping models, LoRAs, or auxiliary ControlNets.
• –Multi-reference conditioning supports up to 16 input images to enforce character, asset, and style consistency across generative runs.
• –ComfyUI support and Hugging Face weights allow immediate self-hosted deployment without third-party API dependencies.