MOSS-Image prompt guide - how to write prompts that work

MOSS-Image reads prompts of up to 4K tokens, so the usual advice for older image models (keep it short, stack keywords, add "masterpiece, 8k") works against it. It does best with a clear, specific description in plain sentences. This guide shows how to write one, with examples you can paste into the studio.

1. Describe the scene in a fixed order

A reliable order is: subject → action → setting → light → camera or medium → mood. Each part answers one question the model would otherwise guess.

An elderly fisherman mending a net on a wooden pier at dawn, weathered hands in sharp focus, soft mist over the harbour behind him, low warm side light, 85mm portrait lens, quiet and unhurried.

Compare that with "old fisherman, pier, dawn, cinematic, 8k". The short version leaves out who does what, where the light comes from and how close the camera is, so every run picks differently.

Say what matters first. If the image is about the product, start with the product. If it is about the headline, start with the headline.

Prefer concrete words to quality words. "Soft north-window light" changes the image; "beautiful lighting" mostly doesn't.

2. Put visible text in quotes

MOSS-Image is trained to render text. Tell it exactly what to write and where:

A hand-painted wooden sign above a mountain bakery door reading "FRESH BREAD DAILY" in cream serif letters, pine trees behind, early morning sun.

Tips for clean lettering:

• Quote the exact wording, with the capitalisation you want.
• Say where it goes: on a sign, at the top of the poster, on the label of the jar.
• Describe the type: serif or sans-serif, hand-painted, neon tubes, embossed.
• Keep it short. One headline and one small line work far more often than a paragraph.
• Force line breaks yourself. The official user guide notes the model wraps text on its own, not always where you want. Write the line breaks into the prompt where you need them.
• Check at full size. Read every character of the result before you use it.

English has the most training data. Chinese text renders well too; write the Chinese characters directly inside the quotes.

3. Choose the aspect ratio before you write

The frame changes the composition, so pick it first and write for it:

Use	Ratio
Instagram or LinkedIn post	4:5 or 1:1
Story, Reel, phone wallpaper	9:16
YouTube thumbnail, slide, blog header	16:9
Poster, book cover	2:3
Cinematic banner	21:9

For a tall frame, describe something tall: a figure, a doorway, a tower. For a wide one, describe a horizon or a scene that unfolds sideways.

4. Styles: say the medium

The studio's style chips add a short description of the medium to your prompt. You can do the same in your own words, and be more specific:

• Photo: lens, distance, light source, time of day, film or digital.
• Anime: line weight, shading (cel or painterly), palette, setting.
• Illustration: medium (watercolour, gouache, ink), texture, era.
• 3D: material (clay, glass, plastic), lighting, background.
• Poster: hierarchy (headline, subline, image), palette, margins.

If you write your own style description, set the style chip to No style so the two don't compete.

5. Tags work too

MOSS-Image also accepts tag sequences. The user guide recommends this order for best results: quality, score, meta, year, character, series, artist, general. For example:

best quality, highres, 1girl, short silver hair, amber eyes, oversized olive hoodie, rainy street, neon reflections

Tags are handy for anime-style images. For anything with text or a precise layout, sentences are clearer.

6. Negative prompts

The base model's user guide lists these as effective negative tags:

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

Hosted services decide whether a negative prompt is sent. Ours keeps the provider's defaults, so you get good results without one; describe what you do want instead of listing what you don't.

7. Iterate one change at a time

1. Generate two to four images from the first prompt.
2. Pick the closest one and change one thing: the light, the framing or the wording.
3. To keep a composition while you adjust words, set a seed under Advanced and reuse it.

Copyable starting points

Product shot

Amber glass perfume bottle standing on a slab of wet slate, a single white orchid petal beside it, soft rim light from behind, deep green backdrop, studio product photograph.

Poster with a headline

Minimal travel poster for a fictional island called "ISLA VERDE", bold sans-serif title across the top, stylised volcano and palm silhouettes, teal and coral palette, wide cream margins.

Anime scene

Anime scene of a girl in a yellow raincoat waiting at a rural bus stop in summer rain, puddles reflecting blue hydrangeas, soft afternoon light, clean line art.

Neon text at night

Neon sign in a narrow night alley spelling "NOODLES 24/7" in pink and blue tubes, steam rising from a food stall, wet cobblestones reflecting the light.

For exact placement of several elements, move on to JSON layout prompts. For the full set of examples, with the images they made, see the gallery.