Guide

How do I write prompts for MOSS-Image?

A member of the MOSS-Image AI product team · 2026-10-26

Describe the scene, quote the exact text, and specify line breaks. The official guide recommends English for scene descriptions; keep visible text in its intended language. Use a JSON layout to guide placement, while allowing for differences between the draft and the generated image.

1. Make the scene and wording specific

MOSS-Image reads prompts of up to about 4,096 tokens. Spell out the subject, setting, style, colors, lighting and the exact copy. In our eight-scene development test, four raw outputs had text problems. Planned prompts improved those examples in our visual review. Six scenes requested visible text; the other two did not, and we did not measure character-level accuracy.

To save time, use MOSS-Image AI to expand your idea and generate an image automatically, then refine the wording or saved layout if needed.

2. Quote the words, control the line breaks

Put visible text in quotes: a sign that reads "GRAND OPENING". The model inserts line breaks on its own, but not always where you want them. For exact breaks, write \n or a real line break inside the quotes.

3. Use a negative prompt to specify what to avoid

A positive prompt describes what you want in the image. A negative prompt describes unwanted features. For example, worst quality in the negative field asks the model to avoid very poor image quality; it does not request a poor-quality image.

The official Base user guide suggests these terms:

nsfw, worst quality, bad anatomy, watermark, jpeg artifact

They refer to explicit adult content, very poor quality, anatomical errors, watermarks, and JPEG compression artifacts. If your workflow exposes a Negative prompt field, enter them there—not in the main prompt. If your interface has no separate negative field, skip this step. These are guidance terms, not a guarantee that every unwanted feature will disappear.

4. Tag prompts for anime

MOSS-Image also understands tag sequences. The recommended order is quality → score → meta → year → character → series → artist → general, for example best quality, highres, 1girl, blue hair.

5. Control positions with a JSON layout

A layout has a summary (overall caption) and items, each with a label (text, object or character), a caption, an optional text and a bbox. Boxes use [y_min, x_min, y_max, x_max] in pixels measured from the canvas center (abs_centralized, yxyx). A 1024×1024 canvas runs from −512 to +512 on both axes.

{"summary":{"meta":["poster"],"canvas":"single","caption":"A minimalist travel poster with a cream background and a blue mountain illustration."},
 "items":[{"label":"text","caption":"Large dark-blue uppercase title.","text":"EXPLORE","bbox":[-416,-384,-288,384]},
          {"label":"object","caption":"A flat blue mountain illustration.","bbox":[-192,-288,384,288]}],
 "bbox_coord_space":"abs_centralized","bbox_format":"yxyx"}

Tip: keep positions in the boxes and out of the captions. “Title at the top” in a caption fights with a box you drag to the bottom.

Skip the JSON

MOSS-Image AI expands your idea into an image, with a saved layout you can review and adjust afterwards.

See the official user guide for settings and formats, and our text rendering tests for observations and original images.