Writing prompts
MOSS-Image was trained on long, careful descriptions of pictures. It does best when you write like that too.
Describe, don’t command
Say what is in the picture, where, and how it looks: subject, setting, composition, colors, light, and style.
| instead of | try |
|---|---|
| a cat | A ginger cat asleep on a windowsill, curled in a patch of afternoon sun, potted herbs beside it, soft focus background, warm film photograph |
| poster for a jazz night | A 1950s-style jazz poster: a silhouetted saxophonist in deep blue on a mustard background, the title “BLUE NOTE NIGHT” in large white letters at the top, “Friday 9 pm” in small letters at the bottom |
English works best; other languages, Chinese in particular, also work.
Text in pictures
MOSS-Image is unusually good at writing. Put the exact words in quotation marks and say where they go. To control line breaks, start new lines in the prompt itself:
A handwritten note on a fridge door:
"Milk
Eggs
Bread"
Negative prompt
Negative prompt lists what you don’t want. MOSS-Image’s authors suggest:
worst quality, bad anatomy, watermark, jpeg artifact
Leave it empty if you don’t need it.
Tags
Short comma-separated tags also work, for example best quality, watercolor, lighthouse, stormy sea, gulls.
Put quality words first and subject words last.
Layouts
To say where things go, use a layout: boxes drawn on the picture, each with a description and, for text, the exact words.
Let the model help
- Enhance prompt (above the prompt box) rewrites your prompt in three lengths. Pick one with Use as prompt; ⌘Z brings your own prompt back.
- Describe image writes a description of any picture, a good starting prompt for something similar. See Describing images.
The character count beside each box is just for reference; MOSS-Image reads up to about 4,000 tokens (roughly 15,000 characters). The limit can be lowered with Max tokens in Advanced options.