MOSS-Image Takes Your Layout as the Prompt: A Free AI Poster Model You Can Use for Client Work

A new open image model reads boxes, coordinates and exact headline text, and its licence lets you bill for the result. Here is how to use it, and where it still falls short.

Every image model has the same habit with posters. You ask for the headline at the top, and it lands in the middle. You ask for the price in the corner, and it appears twice, spelled differently. The usual fix is to keep rerolling and hope, or to give up and set the type yourself afterward.

MOSS-Image, released on October 16 by the OpenMOSS team, takes a different kind of prompt. Alongside an ordinary description, it accepts a small structured file that lists each element of the picture, says what it is, gives the exact words to print, and gives the box it should sit in, measured in pixels. That is not a prompt so much as the notes on a designer's comp. And the licence on the model weights is Apache 2.0, which means you can use what it makes for paying clients without a revenue ceiling.

My position, after reading through its guide and the first two days of community work around it: this is the most designer-shaped open image model released this autumn, and the layout feature is worth learning now, even though its own makers call it experimental. You will lose some afternoons to it. The skill you build transfers.

Why a box beats a paragraph

Describing position in words is a losing game. "Large title centred near the top" can mean a dozen different things, and the model picks one. A box can only mean one thing.

MOSS-Image's user guide shows the idea with a simple travel poster. The picture is described once, as a whole scene. Then it lists two elements. One is a text element: the word "EXPLORE," described as a "large dark-blue uppercase title centered at the top," with a box across the upper part of the frame. The other is an object: "a flat blue mountain illustration in the lower center," with a box below it.

You can mix the two approaches. Write a full description of the mood, palette and style in plain language, then pin down only the pieces that have to land in a particular place. The model also accepts long descriptions, up to 4,000 tokens, which is enough for a proper creative brief rather than a one-line wish.

The guide also supports a third style, a comma-separated list of tags, which will be familiar to anyone who came from anime-style image tools. For poster work, the boxes are the point.

The licence is the other half of the story

Open image models have spent the past few weeks arriving with strings attached. Qwen-Image-2.1 and the add-ons built on it carry a research licence. The InfoVis slide model ships under a non-commercial FLUX licence. FLUX 3 Image Klein lets you use outputs commercially only if your business earns under US$5 million a year.

MOSS-Image's model card says plainly: "The model weights and bundled inference code are released under the Apache License, Version 2.0." For a freelancer or a small studio, that removes the question you would otherwise have to answer before every job. You can make the poster, deliver it, and invoice for it.

This matters more than any quality comparison. A slightly better-looking model you cannot use on paid work is a hobby. A capable model you can use anywhere is a tool.

How the boxes work, in plain terms

This is the part that trips people up, so it is worth slowing down.

Most design apps, including Figma, Photoshop and Illustrator, measure position from the top-left corner of the canvas. MOSS-Image measures from the centre. On a 1024 by 1024 canvas, the middle of the picture is zero, the left and top edges are minus 512, and the right and bottom edges are plus 512.

Each box is four numbers in this order: top, left, bottom, right. (The guide calls this format "yxyx," because it lists the up-down position before the side-to-side one.)

So the guide's headline box, [-416, -384, -288, 384], reads like this:

• Top edge 416 pixels above the centre.
• Left edge 384 pixels left of the centre.
• Bottom edge 288 pixels above the centre.
• Right edge 384 pixels right of the centre.

That is a band 128 pixels tall and 768 pixels wide, near the top of the frame. The mountain's box, [-192, -288, 384, 288], is a larger block starting a little above the middle and running toward the bottom.

To convert from your design app, the guide gives the rule: subtract half the canvas width from every left-right number, and half the canvas height from every up-down number. A headline your comp places 96 pixels from the top on a 1024 canvas sits at minus 416 in MOSS terms.

Here is a skeleton of the structure for a 1024 by 1024 poster, built from the field names in the guide (the wording is my own example):

{
  "summary": {
    "meta": ["poster"],
    "canvas": "single",
    "caption": "A warm screen-printed poster for a night market, two-colour orange and navy, paper grain."
  },
  "items": [
    {
      "label": "text",
      "caption": "Bold condensed navy headline across the top.",
      "text": "NIGHT MARKET",
      "bbox": [-448, -416, -320, 416]
    },
    {
      "label": "object",
      "caption": "Rows of glowing paper lanterns over food stalls.",
      "bbox": [-256, -448, 320, 448]
    },
    {
      "label": "text",
      "caption": "Small navy date line at the bottom.",
      "text": "FRIDAYS 6PM",
      "bbox": [384, -192, 448, 192]
    }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

The whole thing goes in as the prompt, exactly as written.

Put this into practice

Start with the free demo, but know its limit. OpenMOSS runs a hosted demo on shared free hardware. Paste the structure into its prompt box. Expect queues.

Run it yourself on an NVIDIA card through ComfyUI. ComfyUI added built-in support on release day, ahead of its next numbered release, so update to the latest version before you look for the nodes. ComfyUI's repackaged files give you the sizes that decide whether it fits. The picture model is 10.2 GB, or 5.4 GB in a shrunken version. The part that reads your prompt is a separate 9.8 GB or 7.3 GB file. With both shrunken files you are looking at roughly 13 GB on disk. OpenMOSS's own ComfyUI plugin ships a ready-made workflow file, moss-image.json.

Use the default settings first. The guide recommends 50 sampling passes (the number of refinement rounds per image), a guidance setting of 3 (how strictly the model sticks to your prompt) and a "shift" of 5 (a timing setting for those passes; leave it alone). A higher guidance setting can help short prompts; long, detailed prompts usually do fine at 3.

Keep canvas sizes to multiples of 16. The guide suggests about 1, 1.5 or 2 megapixels in total, at any aspect ratio. 1024 by 1024 is a safe first test; 1024 by 1536 suits a portrait poster.

Write the words in English when you can. The guide says English has the most training data. Chinese and other languages are supported, but less reliably. If a line must break in a specific place, put a line break in the text yourself; the guide says automatic breaks "may vary."

Change one box at a time. Fix the seed (the number that controls the random starting point) and move a single box, or change a single line of text, between runs. That tells you whether the model is actually following your layout, or just happening to land near it.

Draw the comp first. If you already sketch thumbnails, you are ahead. Measure three or four key positions on your sketch, convert them, and leave everything else to the scene description. Boxing every element tends to fight the model; boxing the elements that carry meaning is where the gain is.

Where it falls short

The layout feature is experimental, by its makers' own label. The user guide puts it under "Layout Prompts (Experimental)." Expect boxes to be treated as strong suggestions rather than rules, and plan to check every important element.

There are no published quality numbers. OpenMOSS publishes no benchmark scores and no limitations section. The only comparison on the model card is a single before-and-after between its two versions. You are the test.

It is slow. Fifty passes per image means waiting, not rapid rerolling. A community fast version that cuts this to 4 to 8 passes appeared on October 18, but its own card calls it work in progress and warns that converting it to the usual lighter file format "discards most of the distillation," meaning most of its speed advantage.

Results do not match across tools. The ComfyUI plugin's notes say the same seed and settings "do not guarantee identical images" compared with OpenMOSS's standalone app. Pick one tool for a project and stay in it.

It needs an NVIDIA card. The model card asks for an NVIDIA GPU with CUDA. Mac users are limited to the hosted demo for now.

The prompt-writing helper is built for someone else's assistant. OpenMOSS ships a skill that turns a rough idea into a layout file, but the README says it is "designed for GPT-6-family agents." It may work elsewhere. It is not promised.

It is two days old. The supporting tools, from fast versions to style add-ons, appeared within a day of release and none publish measurements yet.

What this asks of you

For three years, the craft of AI image work has been describing things well. MOSS-Image hints at a different craft: placing things well, then describing only what placement cannot say.

That is good news for anyone who already thinks in grids. Take one real brief this week, sketch the comp, measure three positions, and build the layout file by hand. If the headline lands where you drew it, keep going. If it does not, you have lost an afternoon and learned exactly how much the "experimental" label means today. Either way, you will know before a client asks.