Write the Layout, Not the Vibe: What MOSS-Image Gets Right About Poster Prompts

Most poster prompts are a mood with a wish attached. "Bold retro travel poster, warm sunset, big headline at the top, palm trees, 1970s print texture, high detail." Something comes back. It is rarely the thing you pictured, and you cannot say which decision failed, because you did not make decisions. You made suggestions and let the sampler vote on them.

MOSS-Image, a 5B open-weight model from MOSI.AI with SII, Fudan University and Shanghai Jiao Tong University, is worth reading for one reason beyond its scores: the authors treat that failure as two separate problems. Rendering is one. Specification is the other. Their argument is that neither fixes the other. A perfect renderer still guesses your hierarchy. A perfect brief still gets smeared by a model that cannot hold six lines of type.

The fix they built is unglamorous and, for anyone who writes design briefs for a living, immediately familiar.

What the model actually is

A 5B single-stream diffusion transformer, pretrained from scratch rather than fine-tuned off an existing base. Text tokens come from a frozen Qwen3.5-9B encoder, image latents from a frozen FLUX.2 VAE, and the blocks use joint self-attention with a modified feed-forward layer. The paper is a 35 page technical report and the repository is public; the Japanese field report that surfaced it this week describes the release as Apache 2.0, which matters if you ship commercial work.

Two training decisions are more interesting than the architecture.

The first is that they stopped filtering their corpus. After deduplication and removal of AI-generated images they keep roughly three billion images out of five, then deliberately keep the blurry faces, the watermarks, the compression artifacts, and describe them in the caption instead of deleting the image. Their ablation on a filtered model is blunt: overall generation quality got worse, and the OneIG text score collapsed from 0.5243 to 0.1180. They concluded that a model asked to render real screenshots and real packaging needs to have seen ugly real images, with the ugliness labeled.

The second is that natural language and layout JSON enter through the same conditioning pathway. This is the detail to hold onto. Layout control is usually bolted on as a sidecar, a ControlNet or an adapter that injects spatial hints into a frozen backbone. Here the boxes are serialized as text and read by the same encoder that reads your sentence. The model learns one interface with two dialects, prose and coordinates, and can be handed either.

The numbers, and the number that gets oversold

Three results are worth knowing.

On LongLongTextBench, their new benchmark of page-length passages averaging about 1,340 characters, MOSS-Image scores 96.3 (Mean-6) against 64.9 for Qwen Image 2512, a 31 point gap among open models. On the longest English and Chinese conditions it holds 97.1 and 92.4. That Chinese drop is the honest part of the table and the reason to test in your own language before committing.

On 7Bench, which measures whether objects land where the boxes said, it reports 0.8786 mIoU_det and 0.7031 Layout AUC, against 0.4449 and 0.20 for Ideogram 4. Read the subscript: mIoU_det averages overlap only across targets that were detected at all, 445 of 576 in this run. Across every target, including the misses, MOSS-Image's mIoU is 0.6788. The paper also notes that Ideogram's own report lists roughly 0.69 mIoU on 7Bench while their pipeline measures 0.2433, and attributes the gap to checkpoint variant, prompt format and evaluation details. So treat this as a strong signal about where layout adherence lives, not as a clean head-to-head margin.

Then there is RealLayoutArena: 61 design tasks, 20 posters, 20 infographics, 21 interfaces, each posed as a short Chinese request and judged in blind pairwise human comparisons. MOSS-Image lands at 1698.70 Elo, second overall, behind GPT Image 2 and ahead of Seedream 5.0 Pro.

That second place is the number most coverage will repeat, and it is the least useful one. The end-to-end system includes a planner built on GPT-6 that routes the request to task-specific instructions, retrieves reference material, and writes the executable layout. So the arena is measuring a closed commercial model with an open renderer attached, not an open model you can run tonight.

The cleaner measurement is the one buried in the ablation. Hold one enhanced prompt fixed for every generator and compare the backbones, without letting each system add its own prompt magic: the MOSS-Image backbone tops that group at 1807.34 Elo. The planner still wrote the prompt, so this is not a planner-free run. It is the closest thing in the paper to an apples-to-apples renderer comparison, and the open model wins it. Keep that distinction in your head whenever someone quotes the 1698.

The part you can steal today

You do not need the weights to use the finding. The paper's prompt-strategy comparison ran five briefs for the same 61 tasks: the short request, a detailed description, a direct expansion with no agentic routing, the planner's output rewritten as prose, and the planner's output passed as JSON. Their conclusion is that detailed specifications beat concise requests in human preference, by a margin large enough to matter.

That result is not surprising. What is useful is the shape of the specification the planner is required to produce, because it is a better brief template than anything most teams write by hand. Eight root fields: scene type, a global brief, aspect ratio, layout, elements, anchors, must-nots, palette.

Elements carry identity, appearance, literal text, typography, and a box. Anchors describe relative position. Must-nots are explicit exclusions. Boxes are integers in normalized 0 to 1000 coordinates, top-left origin, ordered y-min, x-min, y-max, x-max, with positive width and height, and overlap is allowed when the composition calls for it.

Two constraints in the implementation are worth copying even if you never touch this model.

Text fit is checked in physical pixels. For an output of width W and height H, a box spanning delta-x on the 1000 scale gives delta-x times W over 1000 real pixels. A planner that does not do that arithmetic will happily put a 14 character headline in a 60 pixel slot and call it a layout.

Intended line breaks are escaped into the text field. The paper's case study shows why: with line breaks specified, the character content of each line becomes distinguishable around denoising step 8, and without them several rows stay ambiguous to the final step. Your copy is not a string, it is a stack of lines with a rhythm, and saying so early changes the render.

There is also a quiet instruction in their planner spec that reads like design advice: boxes describe coherent visual units, not mandatory visible panels. A box around a headline and its kicker is one unit, not two rectangles to be outlined.

Where this still breaks

Normalized coordinates have a transfer problem. In the paper's appendix, the same layout generated at a resolution the model never saw during training behaves differently depending on the coordinate convention. Relative 0 to 1000 coordinates pull the composition into a smaller central region, leaving wide margins. Center-relative absolute pixel coordinates hold the spatial scale closer to the original. The model was trained on a mix of normalized, absolute, and center-relative boxes precisely so it would not depend on one, and the ablation still favors the absolute form for resizing.

Practical translation: if you build a layout at one canvas and reuse it at another, expect the margins to drift and budget a pass to fix them.

The second limit is duller. Prompt adherence is not composition. A model that puts every element exactly where you asked will also put a bad hierarchy exactly where you asked, because the specification encoded your taste, including the parts of it that were wrong. The Japanese field report makes the same point from the other side, separating reference use from editing: a vision encoder can pull the atmosphere of a reference image into the plan, but that is not the same operation as editing an existing file, and treating the two as one step is how people end up with drift.

And the practical entry point is gated. The hands-on ComfyUI workflow, including the downloadable graph and notebook, sits behind the author's paid membership. What is public is the paper, the repository, and the shape of the approach, which is enough to change how you write briefs and not quite enough to reproduce her exact pipeline.

A brief you can use this week

Take your next poster prompt and rewrite it as the eight fields, in whatever model you already pay for.

Name the scene type and the aspect ratio first, because those are decisions, not vibes. Write the global brief as one sentence about intent, not a pile of style adjectives. List elements with their literal text and a rough box in 0 to 1000 coordinates. Add two anchors that describe what sits relative to what. Add three must-nots, which is where most slop gets prevented: no floating objects, no gradient mesh, no extra copy. Pick a palette of four values.

Then run the same brief twice, once as prose and once with the boxes included, and compare at full size rather than in a thumbnail grid. If the model you are using ignores coordinates entirely, you have learned something useful about where its prompt adherence actually lives, and you can go back to short prompts without wondering whether you left quality on the table.

The larger point is the one the benchmark table makes without saying it. For two years the poster problem has been framed as a taste problem, and the answer has been "write better prompts." That advice is mostly about adjectives. The evidence here points somewhere less romantic: the models follow geometry when you specify geometry, and the reason AI posters still look like AI posters is that almost nobody specifies any.

If you want the visual side of that argument, the backlash against machine-made flyers is worth reading alongside this. Same artifact, opposite end of the pipeline.

Related reading: why AI posters make people angry, decomposing a composition prompt, and what text rendering actually costs you.