MossImage

The thing that kept breaking for me was not the picture. It was the text.

I wanted a simple diner menu. The image model gave me a beautiful plate of pancakes and a price that read like a captcha. Posters had the same problem: good composition, unusable headline. Text is treated as texture, not content.

MossImage is our attempt to fix that specific job. It runs MOSS-Image SFT, an open-source model built for text rendering and layout. You quote the exact words, and if you need precision, you place them with a JSON layout instead of hoping the prompt lands. We added a prompt library with 38 tested poster, menu, packaging, UI and ad prompts, plus 8 showcase renders on the homepage so you can see the failure cases we care about.

It is early. The MOSS-Image generator is text-to-image only — no image-to-image or inpainting yet — and English is still the most reliable language, with Chinese supported. I would love feedback on text fidelity, dense layouts and which format you would want next.