MOSS-IMAGE · PROMPTS · CHECKED OCTOBER 21, 2026
MOSS-Image follows long prompts closely and can render English and Chinese text, so a good MOSS Image prompt reads like a short brief: what is in the picture, where it sits, how it is lit and which words must appear. Below are the three prompt formats the model accepts, the rules for text and line breaks, the settings that matter, and example prompts you can run right away.
Format rules come from the official MOSS-Image user guide. The example prompts were written by MossImage, not by OpenMOSS.
| Format | Best for | How to write it |
|---|---|---|
| Natural language | Most images: posters, products, scenes | Full sentences. Subject and placement first, then light, materials and style. |
| Tag sequence | Illustration and character art | Comma-separated tags in this order: quality, score, meta, year, character, series, artist, general. |
| JSON layout (experimental) | Exact positions for text and objects | A summary plus a list of items, each with a bounding box. Build one with the layout tool. |
You can also mix them: describe the scene in a sentence, then list the exact elements. English gives the most reliable results because it has the most training data, but Chinese and other languages are supported.
Text rendering is one of MOSS-Image's strengths, especially for posters and banners. A few habits make it more reliable:
\n.Results vary from run to run. Proofread names, prices and dates in every image before you publish it.
| Setting | Recommended | Why |
|---|---|---|
| CFG | 3 (range 3–6) | Long prompts need little guidance; short prompts may need more. |
| Steps | 50 or more | Fewer steps lose detail unless you use a speed LoRA. |
| Size | About 1–2 megapixels | Any ratio; width and height must be multiples of 16. |
| Negative prompt | nsfw, worst quality, bad anatomy, watermark, jpeg artifact | The short list from the user guide. |
| Sampler | Euler, shift 5 | The defaults in the official pipeline. |
On MossImage you only choose the aspect ratio and the 0.5K, 1K or 2K size; the sampling settings are handled for you. Running it yourself? See the ComfyUI guide.
Copy a prompt, or open it in the editor with MOSS-Image selected and change the subject and the words.
A vintage travel poster of Kyoto at dusk: a red pagoda against an orange sky, flat screen-print style with soft grain. The headline “KYOTO” in large cream serif letters across the top, and “Autumn 2026” in small caps at the bottom.
Quote the headline and say where it goes.
Use this promptA wide sale banner for a night market, neon lights and steam from food stalls, deep blue background. Large text on the left: 夜市狂欢 NIGHT MARKET Smaller line below: 10/24 – 10/26.
Line breaks in the prompt become line breaks in the text.
Use this promptA studio product photo of a frosted glass skincare jar with a gold lid, centered on a pale travertine block, soft window light from the left, gentle shadow to the right, plain warm-white background, no text.
Subject and placement first, then light and materials.
Use this promptmasterpiece, best quality, illustration, 2026, 1girl, original, flat colors, city street, rain, umbrella, neon reflections, looking back
Quality tags first, general description tags last.
Use this promptA chalkboard café menu on a brick wall, hand-lettered in white chalk. Title “今日咖啡” at the top, then three lines: 美式 18, 拿铁 22, 燕麦拿铁 25. Small coffee cup doodles in the corners.
Keep each line short so it can be checked.
Use this promptA clean poster for a running club. Background: sunrise over a coastal road. Elements: a runner silhouette in the lower third; headline “RUN AT DAWN” top center in bold sans-serif; “Every Saturday 6:30” under the headline; a small logo area in the bottom-right corner.
Describe the scene, then list exact elements.
Use this promptWrite your own or paste one from above. The cost appears on the button before anything runs.
Natural language works for most images. The official user guide also supports tag sequences and an experimental JSON layout, and you can mix natural language with structure. English is recommended because it has the most training data, but Chinese and other languages work too.
Put the exact words in quotation marks and say where they go. MOSS-Image can render English and Chinese and adds line breaks on its own, but results vary; for exact line breaks, type the breaks into the prompt. Always proofread the result.
The text encoder reads up to 4,096 tokens, so long, detailed prompts are fine. The user guide says CFG 3 is usually enough for long prompts, while short prompts may need a higher CFG, within the recommended 3 to 6.
The user guide recommends a short tag list: nsfw, worst quality, bad anatomy, watermark, jpeg artifact. Add your own terms only when a specific problem keeps appearing.
About 1, 1.5 or 2 megapixels. The model can go from roughly 512×512 up to 2K, in any aspect ratio, as long as width and height are multiples of 16. On MossImage you pick the ratio and the 0.5K, 1K or 2K size from the menus.
Yes. Each example has a Use this prompt button that opens the MossImage editor with MOSS-Image selected. A 1K image costs 10 credits, and new accounts get 20 free credits.