MOSS-Image Technical Report

MOSS-Image Technical Report. By MOSI. AI, from MOSI Intelligence, the Shanghai Innovation Institute, and two other institutions. Published on the sixteenth of October. Its code repository is public and real: eighteen source files.

Recent closed-source image generators can produce a wide range of visual content, from simple compositions all the way to information-dense graphics. Open-weight text-to-image models still struggle at the dense end of that range. Pages with a lot of text on them, and intricate layouts, are where they fall short. That gap between closed and open systems is what this report sets out to narrow.

MOSS-Image is an open generation system with two halves. One supplies high-fidelity visual synthesis: the ability to draw well. The other supplies intent-aware composition, working out what the user actually wants on the page before anything is drawn. Neither half alone is the claim. Together, they are meant to yield images that are rich in text and coherent in layout.

The synthesis half is a five billion parameter single-stream diffusion transformer, pretrained from scratch with a carefully designed recipe. That recipe is credited with enabling bilingual long-form text rendering and precise layout control. The composition half is an intention-aware agentic prompt enhancement module. It takes diverse user requests, which may be short or loosely phrased, and translates them into detailed specifications of content and layout that the generator can then follow.

To measure this, the authors introduce two benchmarks of their own. LongLongTextBench evaluates long-form text rendering in English and Chinese. RealLayoutArena evaluates layout-rich generation end to end. On LongLongTextBench, MOSS-Image reaches state-of-the-art performance among open-weight models, with an average text coverage of ninety-six point three percent.

RealLayoutArena gives a more layered picture. As a complete system, MOSS-Image ranks second in the end-to-end comparison. When every generator is handed the same enhanced prompts, though, its backbone ranks first. That split is informative, because it pulls the two halves of the design apart. Holding the prompts fixed, the generator comes out on top of those evaluated. Taken as a whole pipeline, the system places just behind the leader.

Open-weight is the operative word throughout. The authors say they will release the model, the training recipe and the evaluation framework, so that others can build on dense-text and layout-heavy generation without relying on a closed system to do it.

MOSS-Image Technical Report. By MOSI. AI, from MOSI Intelligence, the Shanghai Innovation Institute, and two other institutions. Published on the sixteenth of October. This is the long version: the whole paper rather than the abstract.

Ask a strong open-weight image model for an English koala infographic and you get something that looks like an infographic from across the room. Up close, the small text is unreadable, the headings are garbled, and the content can simply be wrong: the paper's opening figure shows one output stating a wrong birth weight. The authors split the problem in two. Rendering is reproducing long text accurately and obeying spatial constraints. Specification is turning a request of a few words into actual content and an actual layout. Neither is enough alone. MOSS-Image is an attempt to build both halves and join them through a single interface.

The renderer is a five billion parameter diffusion transformer, trained from scratch. Single-stream means text tokens and image tokens pass through the same blocks under joint self-attention, rather than through separate towers. The text comes from a frozen Qwen3.5-9B encoder and the image latents from a frozen FLUX.2 autoencoder, and the five billion counts only the generator between them. The design decision everything rests on is this: a plain description and an explicit layout, written as JSON with bounding boxes, typography and literal strings, travel through the same text-conditioning pathway. There is no separate layout module. Layout is just more text, and the model learns to read it.

The data recipe has a contrarian centre. After deduplication and removal of AI-generated images, the corpus falls from about five billion images to about three billion, and nothing else is filtered out. Watermarks, blur and compression artifacts are written into the captions instead. The reason is a preliminary experiment on a smaller model. A conventional filter on fields like brightness, entropy and watermark area kept forty-three percent of the data, and the filtered model got worse. Its OneIG text score fell from about point five two to about point one two. The slogan is annotate rather than filter.

Captions come at three granularities from Gemma 4: a sentence, a paragraph, and a long description, averaging about twenty-one, seventy-seven and two hundred and thirty-five tokens. Visible text is transcribed verbatim with its line breaks kept, because line breaks are a form of layout. A transcription says which words appear; the breaks say how they sit. In a denoising case study, explicit breaks let the content of each line become recognisable early, by the eighth of thirty steps, while without them some rows stay ambiguous to the end.

Training is progressive, raising resolution and text length in stages until a sample can carry four thousand and ninety-six text tokens.

The second half is the prompt enhancer, run on GPT-6 at inference time. It routes a request to poster, infographic or user-interface guidance, searches for knowledge and reference images, and writes a JSON specification of literal text, boxes and typography, checking that text fits the physical size of each box before handing it to the renderer.

On rendering, the authors argue that the existing LongTextBench is saturated: the strong models listed score between ninety-six and ninety-nine, on targets averaging about a hundred and twenty characters. Their LongLongTextBench averages about one thousand three hundred and forty characters per target, roughly eleven times longer, in English and Chinese. MOSS-Image scores ninety-six point three, averaged over six language and length groups. The next open-weight model, Qwen Image 2512, scores sixty-four point nine. GPT Image 2 still leads overall. On the 7Bench layout test, box overlap among detected objects is point eight eight, double Ideogram 4.

The comparison that tests the thesis is the prompt-strategy ablation on RealLayoutArena: sixty-one design requests, judged blind and in pairs by people. A short request alone rates about one thousand and eighty Elo. Every form of specification lifts the same generator above fifteen hundred. The JSON layout tops the group at about one thousand six hundred and twenty-three. The same plan rewritten as prose gets about one thousand five hundred and six, slightly below a direct expansion that used no routing or retrieval at all. In the full system ranking, MOSS-Image is second at about one thousand six hundred and ninety-nine, behind GPT Image 2. Given identical enhanced prompts, its backbone ranks first of nine generators.

There is a short limitations section. The backbone cannot condition on reference images or edit images, and the prompt enhancer depends on external language models whose aesthetic judgment, reading of references and factual errors can degrade the specification however faithfully it is rendered.

The appendices add more. The prompt ablation and the shared-prompt comparison used an earlier planner that handled all sixty-one tasks in one shared context, not the final per-task version. The shared-prompt ranking rests on five hundred and seventy-six votes covering about a fifth of task pairs, with intervals that do not adjust for correlation within a task or a rater. The text metric ignores word order. And training captions are predominantly English, without dedicated Chinese captions: the longest Chinese passages are the weakest group, at ninety-two point four.

Two ideas carry away. The first is that layout can simply be text: if the renderer is trained on boxes that match real images, a planner can steer it with nothing more than a carefully written prompt, and the structured form of that prompt beats the same plan in prose. The second is the data lesson. Filtering for clean-looking images cut the text score to under a quarter of what the unfiltered data gave. Keep the messy data, and describe the mess.