MOSS-Image Prompt Generator

Free · No sign-up · Runs entirely in your browser

The MOSS-Image Prompt Generator builds prompts for MOSS-Image, a text-to-image model from the OpenMOSS Team: a 5B DiT with a Qwen3.5-9B text encoder and FLUX.2 VAE. It understands three prompt styles in one model — natural language, tag sequences, and JSON layout for precise text placement and object arrangement. Describe your scene, pick a style, or build a full layout — everything runs locally in your browser; nothing is uploaded.

Go to Generator

Generate Prompt

Put rendered text in quotes and line breaks for best results.
Output: comma-joined non-empty tags.
Coordinates are absolute pixels from the canvas center, in yxyx order — the position presets compute them for you. Items with an empty caption are skipped.

Generated Prompt

Your generated prompt will appear here

Preset Examples

Travel Poster (Layout)

Layout mode: summary + 2 items with captions and positions

Coffee Shop Sign (Natural)

Natural mode: scene, photorealistic, render text "OPEN"

Anime Character (Tags)

Tags mode: subject + style + quality tags

How to Write Better MOSS-Image Prompts

  1. Put rendered text in quotes, and use line breaks inside long text. Both act as layout cues — quotes mark the exact words, line breaks tell the model how to split them.
  2. English prompts recommended; Chinese supported. Clear descriptive language produces reliable results.
  3. JSON coordinates are centered yxyx. Presets compute them for you; on a 1024 canvas each axis runs -512 to +512 from the center.
  4. SFT checkpoint for generation; Base for fine-tuning/LoRA. Use the appropriate model version for your task.
  5. Output dimensions must be multiples of 16. Supported range is 512px to 2K, at any aspect ratio.
  6. Default settings: 50 steps, CFG 3, Flow-matching Euler, shift 5. Raise CFG for short prompts so the model follows them more closely.

FAQ

What is MOSS-Image?

MOSS-Image is a text-to-image model from the OpenMOSS Team — a 5B DiT with a Qwen3.5-9B text encoder and FLUX.2 VAE, Apache-2.0 licensed. It supports natural language, tags, and JSON layout prompts in a single model.

Base vs SFT?

Base = for fine-tuning/LoRA training. SFT = supervised fine-tune, recommended for generation.

How does JSON layout work?

A JSON layout has a summary (scene caption) plus an items array. Each item has a label (text or object), a caption, and a bounding box in centered yxyx coordinates — [y_min, x_min, y_max, x_max], with the canvas center as origin. Text items also carry the exact words to render.

Which prompt style should I use?

Natural language for scenes, tags for anime-style control, JSON layout for posters with positioned text and objects.

Where can I run MOSS-Image?

Hugging Face diffusers pipeline, native ComfyUI, and Civitai (quantized versions).

Affiliated with OpenMOSS?

No. This is an unofficial fan-made tool, not affiliated with OpenMOSS.

Copied!