An image prompt often mixes three jobs: describing a scene, specifying the words that should appear, and deciding where everything belongs. MOSS-Image is interesting because its published interface gives the third job a structured form. This guide separates the documented model features from our own design suggestions. We have not run model inference for this website, and the posters shown here are authored SVG mockups, not evidence of generation quality.

Text is part of the composition

The model’s documentation demonstrates Chinese and English text inside images. Its user guide recommends English prompts, which it says have the most training data, and states that Chinese and other languages are also supported. That is relevant to posters, labels, menus, and diagrams, where language carries information rather than merely suggesting a mood. The user guide also notes that automatic line breaks may differ from what the author intended. Its recommendation to specify line breaks is a useful reminder: the more important a piece of copy is, the more explicitly you should describe it.

Our practical suggestion is to separate exact text from surrounding art direction. Write the intended title once, preserve its punctuation, and state which language belongs to each line. Then describe its visual treatment. Do not use a generated preview as the final authority for a price, address, or event date. Read every character before publication, and keep an editable copy of the original wording outside the image.

A layout is data, not a guarantee

The official layout format uses a scene summary and an array of items. An item can describe text or an object, with a bounding box indicating its intended location. The following is an original teaching example using the documented field names. It describes a square poster with a bilingual heading and an abstract green shape. It is not a saved model result or a promise that every run will follow each box perfectly.

{
  "summary": {
    "meta": ["poster"],
    "canvas": "single",
    "caption": "A cream poster with dark green type and an abstract green circle."
  },
  "items": [
    {
      "label": "text",
      "caption": "Large heading, centered in two lines",
      "text": "SLOW DAYS\n慢日子",
      "bbox": [-420, -420, -120, 420]
    },
    {
      "label": "object",
      "caption": "A single green circle",
      "bbox": [-20, -280, 420, 280]
    }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

For this 1024 by 1024 canvas, the center is the origin. Each axis extends from minus 512 to plus 512. The four values are ordered as minimum y, minimum x, maximum y, maximum x. These are centered pixel coordinates, not percentages and not normalized values from zero to a thousand. Keeping that distinction visible prevents a seemingly reasonable layout from describing the wrong part of the canvas.

The project guide calls layout prompting experimental. Treat the boxes as a way to communicate composition and compare variations. Our interactive homepage demonstrates how changing those boxes changes a browser-rendered layout; it does not call a model. That limited demonstration is deliberate: it teaches the input mechanism without presenting a deterministic SVG renderer as evidence of deterministic image generation.

Long prompts need hierarchy

The project advertises prompts of up to roughly four thousand tokens; its ComfyUI plugin documents a 4096-token text-encoding cap. Tokens are not the same as words, especially across languages. This gives room for a detailed brief, but it is not an invitation to repeat an instruction in five different ways. Conflicting descriptions make a brief harder to interpret, even when it fits within the context limit.

Start with the composition, follow with exact copy, then describe materials and lighting. Keep secondary constraints together. If the text must be legible, describe the space reserved for it instead of allowing decorative objects to compete with the same area. This hierarchy is editorial advice rather than a documented benchmark advantage. Evaluate it against your own examples and retain the version that improves the result you actually need.

Base or SFT?

The model cards describe SFT as a curated-data fine-tune of Base. For an initial creative evaluation, SFT is a reasonable first candidate. Base is relevant when the goal is adaptation or studying the underlying model. Avoid translating those roles into absolute claims such as “SFT is always better” or “Base cannot make finished images.” The right comparison uses the same brief, records settings, and checks the specific qualities your project values.

Fine-tuning also requires more than downloading a checkpoint. You need appropriate data, a repeatable evaluation set, and a way to compare a changed model with its starting point. Keep a small set of difficult bilingual examples separate from your training material. Otherwise, an attractive demonstration may tell you very little about whether the adaptation generalizes to the next brief.

Try the documented ComfyUI plugin

The project’s ComfyUI README describes a custom-node installation, while its main README also points to ComfyUI Manager and says integration into the official ComfyUI repository is still being worked on, so do not expect it to be built in. For the manual path, copy the repository’s comfyui/moss-image/ directory into ComfyUI/custom_nodes/moss-image/ and restart ComfyUI. Import the supplied example workflow by dragging its JSON file into the interface. These steps come from OpenMOSS’s plugin documentation; we did not verify a separate MOSS-Image tutorial on the ComfyUI documentation website.

Download the complete Base model directory linked by that guide. The “5B” in the model’s description refers to the diffusion transformer backbone; the full pipeline also loads a separate text encoder and VAE, so plan storage and memory from the official instructions rather than from the parameter count. In MOSS-Image HF Loader, set model_path to the local directory containing the transformer, text encoder, tokenizer, and VAE folders. The loader expects local files, not a Hub repository ID. Edit the prompts and click Run. Follow the current upstream guide for environment requirements and sampling settings; this marketing site does not install ComfyUI or download model weights for you.

Keep the first experiment modest: one title, one subject, one composition. Save the checkpoint revision and settings alongside the output. Only after that baseline works should you increase text density or change runtime details. The most useful outcome is an understandable experiment, including its mistakes, rather than a single attractive image that cannot be reproduced.

Next steps: read how to write a poster brief, place this release in context with the history of the MOSS family, or check every claim above on our references page.

Sources

Primary sources checked October 18, 2026. This is independent editorial commentary, not an official project publication. See the complete reference and fact-check list.