Anyone who has asked an image model to design a poster knows the failure mode. The illustration looks fine, and then the headline comes out as scrambled pseudo-letters, a paragraph gets cut off mid-sentence, or the price tag drifts into the product photo. For English this is annoying. For Chinese, where a single wrong stroke changes the meaning of a word, it has made text-heavy design work practically off-limits for open image models.
A team in Shanghai thinks it has a fix worth shipping as weights anyone can download. On October 16, the OpenMOSS collaboration — the group behind the original MOSS chatbot, now spanning Fudan University's NLP lab, Shanghai Innovation Institute (上海创智学院), Shanghai Jiao Tong University and the company MOSI.AI (模思智能) — released MOSS-Image, a text-to-image (文生图) model built from scratch around two unglamorous but commercially decisive skills: rendering long passages of Chinese and English text accurately, and placing every element where a layout says it should go.
Key takeaways
- MOSS-Image is a 5B-parameter, single-stream diffusion transformer (DiT) released as two checkpoints, Base and SFT, under the permissive Apache-2.0 licence.
- It accepts prompts of up to 4K tokens and can take a JSON layout that specifies the position and content of every text block and object on the canvas.
- On the team's own long-text benchmark, it reports 96.3% average text coverage, the best among open-weight (开源权重) models; an independent third-party comparison has not yet been published.
- It runs on consumer hardware and plugs into tools designers already use: ComfyUI, Diffusers, Gradio, SGLang and vLLM-Omni.
Why text is the wall
Rendering text inside images is not a niche complaint. Posters, e-commerce banners, infographics, event tickets and app screenshots all carry words, and words are where image models fall apart. The model has to translate an abstract description into precise glyph shapes, keep dozens of characters aligned along a baseline, and respect line breaks and paragraph structure. Diffusion models, which build an image by iteratively denoising noise, are naturally bad at this kind of exact, symbol-level fidelity — and worse in Chinese, where thousands of distinct characters must each be drawn correctly.
MOSS-Image attacks the problem at the foundation rather than bolting on a fix. The team pre-trained the 5B backbone from scratch on roughly 3 billion images (a vendor-reported figure), mixing ordinary captions with text supervision and structured bounding-box labels, so the model learns from the start that some regions of an image are, essentially, for spelling. A Qwen3.5-9B text encoder reads the prompt, and a FLUX.2 VAE handles image decoding. The training recipe uses Muon–AdamW optimisation and, notably, was carried out on heterogeneous alternative GPU platforms — the team has not named the chips, a detail that matters in a year when compute supply chains are under strain.
In its technical report (arXiv 2610.23374), the team introduces two evaluation suites of its own: LongLongTextBench (LLTB), which tests long-form English and Chinese rendering, and RealLayoutArena (RLA), which tests end-to-end generation of layout-heavy designs. On LLTB, MOSS-Image reports 96.3% average text coverage, which the team says is state of the art among open-weight models. On RLA, the backbone ranks first among generators evaluated with shared enhanced prompts, while the full system ranks second end to end. A third vendor-reported score, a bounding-box mIoU of 0.8786 on structured layout tasks, describes how closely the drawn elements land on their planned positions. All of these numbers come from the team itself; the weights are public, so independent replication is now possible, but it has not happened yet.
Layout as a first-class input
The more unusual design choice is that layout is a native input format, not an afterthought. A user can hand the model a JSON document that lists each element — a title, an illustration, a caption — with its text and a bounding box, and the model draws the scene to match. Editing works the same way: move a box, change a line of text, regenerate. The coordinates are specified relative to the canvas centre in a documented format, which makes the workflow reproducible rather than a matter of luck.
Because prompting a layout by hand is tedious, the team also ships an intent-aware prompt-enhancement (PE) module that expands a short request into a full specification with font hierarchy, colour palette and coordinates. It is an agentic component that runs outside the image model, so the image model itself stays small and self-contained.
For a global audience, the interesting part is not the leaderboard but the division of labour this represents. Chinese open-weight releases have mostly chased general capability — bigger mixtures-of-experts (混合专家), longer context, higher benchmark scores. MOSS-Image instead picks one painful, revenue-relevant job and makes it an open default. According to Li An, chief scientist at BrainNet (脑机网), the next phase of the open-model race will be won by teams that ship narrow, verifiable competence rather than another generalist checkpoint. A 5B model that renders a bilingual product poster correctly is exactly that kind of artefact.
Running it yourself
The practical deployment story is unusually complete for a two-week-old release. Weights are on Hugging Face and ModelScope under Apache-2.0, which permits commercial use and fine-tuning without negotiation. MOSS-Image had day-0 support in ComfyUI, SGLang and vLLM-Omni, works with the Diffusers pipeline in Python, and ships a Gradio app with memory modes that step components between CPU and GPU. In the team's own low-memory measurement, generating one 1024×1024 image at 50 steps peaked at under 4 GiB of GPU memory; community training tools can start a LoRA fine-tune with as little as 6 GiB. An API through the MOSI platform followed on October 20 for teams that prefer not to self-host.
Honest limitations
- Every benchmark number above is vendor-reported. The checkpoints are public, but no independent evaluation has yet confirmed the 96.3% coverage figure or the layout scores.
- The backbone is a pure text-to-image model. It cannot take a reference image or perform native image editing; structured re-layout is the intended editing path, and the model card still labels layout prompting experimental, with output dimensions required to be multiples of 16.
- The prompt enhancer is an external dependency that relies on a separate language model, and the project's own tooling assumes a fairly capable agent stack.
- Training on unnamed alternative GPU platforms is a statement of resilience, not of measured performance; the team has not published chip-level throughput data.
- At 5B parameters it will not match the largest closed systems on photographic quality or complex poses; its edge is deliberately narrow.
What readers can do now
- If you make posters, infographics or e-commerce assets, try the SFT checkpoint in ComfyUI with a simple JSON layout before judging it from prompt-only tests — layout input is where it differs most from other open models.
- If you maintain a multilingual design pipeline, benchmark LLTB-style prompts in your own languages; Chinese and English are covered, other scripts are untested.
- If you fine-tune open image models, start from Base rather than SFT: the team ships it specifically so the community can train LoRA adapters on top.