MOSS Image AI Generator

MOSS Image, which its authors write with a hyphen, is an open-source text-to-image model released by OpenMOSS on October 16, 2026 under the Apache-2.0 license. It has 5 billion parameters in a single-stream diffusion transformer, uses a Qwen3.5-9B text encoder and the FLUX.2 VAE, and was trained on roughly 3 billion images. It was not built to win at photorealism. MOSS Image exists to do two jobs most generators do badly: render long passages of English and Chinese text correctly, and place elements where you ask.

Why does text in AI images usually break?

Diffusion models learn from captions, and captions rarely spell out every word on a sign. The model learns what text looks like, not what it says, so it produces letter-shaped noise, drops characters or paraphrases your headline. Every extra line is another chance to drift. Chinese suffers most: training captions are overwhelmingly English, so most models never learn how characters are built.

How does MOSS Image handle long text?

The technical report (arXiv 2610.23374) introduces LongLongTextBench, 240 prompts, half English and half Chinese, where the target text averages about 1,340 characters. MOSS Image scores 96.3 on average, 97.1 on long English and 92.4 on long Chinese. Qwen Image 2512 reaches 64.9 and Ideogram 4 reaches 63.5. The prompt window is 4,096 tokens, and a line break in the prompt becomes a line break in the image, which is how you decide where a paragraph wraps.

What is MOSS Image layout JSON?

Instead of writing "title at the top, logo at the bottom left", you can hand the model a JSON object. Each item has a label (text, character or object), a caption, an optional text string and a bounding box in absolute pixel coordinates centered on the canvas. On the 7Bench layout benchmark this scores 0.88 mean IoU, against 0.44 for Ideogram 4 and 0.14 to 0.17 for GPT Image 2, Nano Banana Pro and Seedream 5.0 Pro. Human raters also preferred JSON prompts to detailed prose, 1,622.87 Elo to 1,509.21. It is marked experimental; the format is documented in our MOSS Image layout JSON guide.

Who is MOSS Image for?

Anyone whose images carry words: event organizers making posters, marketers producing infographics and quote cards, small brands that need a label, and bilingual teams working in Chinese and English. ComfyUI, Diffusers, SGLang and vLLM-Omni supported MOSS Image on day one, and the community published more than 30 LoRAs, fine-tunes and quantized checkpoints within four days of release.

How do you write a good MOSS Image prompt?

• Put every string you want rendered in quotation marks; unquoted phrases get paraphrased.
• Say where each line goes and how large it is: "large uppercase title at the top, small date line at the bottom".
• Write Chinese text as Chinese characters, not as an English description of them.
• For exact placement, switch from prose to layout JSON.

More examples, including Chinese ones, are in the MOSS Image prompt library.

How does MossImage fit in?

MossImage is an independent generator and knowledge base built around the model; we are not affiliated with OpenMOSS. MOSS Image has no official hosted API yet, so our online generator is in preview while we finish the engine integration, and every result shows which engine produced it. Layout Mode, for drawing boxes and exporting layout JSON instead of typing coordinates, is in progress. The MOSS Image model guide covers the architecture, recommended settings (50+ steps, CFG 3, shift 5) and running the weights yourself.

What are the limitations?

MOSS Image is text-to-image only; reference images and editing are listed as future work. Long Chinese text trails English, 92.4 to 97.1. The best layouts in the paper relied on an external language model to expand prompts, and layout JSON is still experimental. It needs 50 or more sampling steps, so it is not fast, and the authors are candid that MOSS Image sells text and structure, not raw realism. Proofread every generation; text rendering varies between runs.