MOSS-Image is OpenMOSS's open image model that paints from a sentence or from a JSON note of where each word and object should sit.
One prompt box for words and for boxes
There are two checkpoints, MOSS-Image-Base and MOSS-Image-SFT. SFT is Base trained further on curated data. They share one sampling interface. On the project page, one whiteboard example looks more orderly with SFT. That is their example, not a rule that SFT always writes cleaner text.
- A normal prompt is a sentence. The quick start uses a refrigerator with a Chinese sticky note that lists milk, eggs and bread. The documented canvas for that call is 1456 by 1456.
- A layout prompt is JSON. Their travel-poster example asks for a cream background, the word EXPLORE in a box near the top, and a flat blue mountain in a box below. Boxes are measured from the center of the canvas, in the order top, left, bottom, right. The diagram on the site draws those input boxes. It is not the picture the model returned.
- The project page's recommended sampler is flow-matching Euler, 50 steps, guidance scale 3, shift 5, and skipping the negative branch at layer 22. The Python example leaves the negative prompt empty.
- They say a prompt can run to 4K tokens, and that the backbone is a 5B single-stream diffusion transformer trained from scratch. Natural language and the JSON layout go through the same text path.
- Changing a box, a caption or the words inside it is how the page describes a structured edit. It is still a new generation from that spec, not a brush on a photo you already have.
Two scores, and one of them borrows GPT-6
The technical report is from MOSI Intelligence with SII, Fudan University and Shanghai Jiao Tong University. The numbers below are theirs. This page did not generate an image or rescore a benchmark.
- LongLongTextBench is a test they built: 240 tasks, half English and half Chinese, aimed at page-length text. The backbone, without their prompt enhancer, scores 96.3 on the six-way average they call Mean-6. In the same table Qwen Image 2512 is 64.9. The longest Chinese slice is 92.4, lower than the English long slice at 97.1.
- On their layout metrics the backbone is at 0.88 mean IoU and 0.70 Layout AUC. The same table lists Ideogram 4 at 0.44 and 0.20.
- RealLayoutArena is 61 short design requests: 20 posters, 20 infographics and 21 interfaces. The full MOSS-Image system scores 1698.70 Elo and places second, behind GPT Image 2 at 1844.35 and ahead of Seedream 5.0 Pro at 1661.16. That full system includes a prompt enhancer. The report says the layout enhancer uses GPT-6. The open weights do not include GPT-6.
- When every generator is given the same enhanced prompts, they rank the MOSS-Image backbone first, at 1807.3. That is a different comparison from the end-to-end ranking.
- On the training set, they say deduplication and removal of AI-made images took the pool from about 5 billion images to about 3 billion. After that they keep the remaining pictures, including watermarks and blur, and write those flaws into the caption instead of throwing the picture out.
What you download, and what you still have to bring
- The weights are Apache-2.0, on Hugging Face and ModelScope. The project page says the pipeline code and third-party pieces keep their own licenses. The repository says the standalone pipeline itself is Apache-2.0.
- The documented local path is Python 3.10 or newer, an NVIDIA GPU, and Diffusers with remote code enabled. They do not publish a minimum amount of video memory for a normal run. The Gradio app has a Sequential memory mode, and they say to use that or a smaller image if memory runs out.
- A demo space is on Hugging Face. On 20 October 2026 the repository added a link to the MOSI API platform. This listing does not copy an API price, because the project page does not show one.
- ComfyUI can load a native plugin from ComfyUI Manager. They say folding it into the official ComfyUI repository is still in progress. The release notes also point at day-one recipes for SGLang and vLLM-Omni.
- KohakuMossTrainer is a separate project, announced in the MOSS-Image notes on 18 October 2026. It says adapter training can fit in 6GB of video memory. That figure is about training an adapter, not about generating a picture, and it is that trainer's claim.
- A prompt-enhancer skill in the repo only turns an idea into a caption or layout JSON. It does not call an image service. The automatic enhancer in the paper is the part that calls GPT-6.
The project page is openmoss.ai/MOSS-Image, the code is OpenMOSS/MOSS-Image, and the report is the MOSS-Image technical report. The image on this page is their announcement collage. The sentence printed on it, that the whole picture including that sentence was generated in one pass, is their caption. It is not a Seailife generation.