During a Reddit AMA, MiniMax's H3 team revealed it's working on a dedicated image model — and plans to open its weights. The model will pack text-to-image generation and general image editing into a single system, and it's already in the post-training optimization phase.
The image model shares the same architectural lineage as H3. It will reuse H3's VAE encoder and pair it with a VAE decoder designed specifically for image generation. MiniMax's intended workflow would have the image model generate a first frame, then hand things off to H3 to continue generating video.
The team also noted that H3 itself has already shown glimmers of image generation and editing ability. They only trained the model to predict a final frame from a 'first frame + text description' — there was no dedicated image-editing training. Still, H3 scored well on multiple image-editing benchmarks in zero-shot settings. That's a big reason MiniMax is pushing the same architecture into a full image model.
https://www.reddit.com/r/StableDiffusion/s/jIc8SXhSGt