Skip to content

MOSS-Image

View markdown Try it in the explorer

MOSS-Image is OpenMOSS’s open-weight text-to-image model. It renders long passages of English and Chinese text, follows detailed natural-language prompts and tag sequences, and accepts JSON layouts that place text and objects at exact pixel positions. Prodia serves the supervised fine-tuned release, MOSS-Image-SFT.

MOSS-Image is a 5B-parameter single-stream diffusion transformer. Prompts are encoded by a Qwen3.5 language model and refined inside the transformer, and a FLUX.2 small VAE decoder produces the final image.

Scenario Job type
Notes, signs, posters and whiteboards with long rendered text inference.moss-image.sft.txt2img.v1
Chinese text rendering inference.moss-image.sft.txt2img.v1
Exact placement of text and objects with a JSON layout inference.moss-image.sft.txt2img.v1
Exact output sizes from 512 to 2048 pixels per side inference.moss-image.sft.txt2img.v1

For sub-second generation, consider FLUX.2 Klein. For images above 4 MP, use MiniMax H3 Image.

Job type Description ETA
inference.moss-image.sft.txt2img.v1 Generate an image from a text prompt or JSON layout ~5s at 1024×1024, ~25s at 2048×2048
  • prompt (required) — natural language, a tag sequence or a JSON layout, 1–16,384 characters. The model reads the first 4096 tokens (about 16,000 characters of English, fewer for Chinese); the rest is cut off.
  • negative_prompt — what to avoid, up to 16,384 characters. Empty by default.
  • width, height — output size in pixels, each 512–2048, default 1024×1024.
  • guidance_scale — 1–10, default 3. 3 suits long prompts; short prompts can benefit from up to 6.
  • seed — integer 0–2147483647 for reproducible results.
  • Quote the text to render: put the exact words in quotes and add line breaks (\n) where the text should break.
  • Write in English for best results: English has the most training data; Chinese and other languages are supported.
  • Describe the whole scene: long, complete descriptions work better than short requests. For short prompts, try a higher guidance_scale.
  • Use a negative prompt for clean results: worst quality, bad anatomy, watermark, jpeg artifact is a good start.

A layout prompt is a JSON object with a summary.caption for the whole scene and items, each with a label (text or object), a caption, the text to render and a bbox. Boxes are [y_min, x_min, y_max, x_max] in pixels with the image center as the origin ("bbox_format": "yxyx", "bbox_coord_space": "abs_centralized"), so on a 1024×1024 image both axes run from -512 to 512. Send the layout as the prompt string, with the width and height it was drawn for.

A handwritten note:

{
"type": "inference.moss-image.sft.txt2img.v1",
"config": {
"prompt": "A wide view of a refrigerator beside a kitchen counter with a fruit bowl. A small sticky note on the refrigerator door is handwritten: \"Buy milk and eggs!\"",
"seed": 42
}
}

A poster from a layout (the prompt is the layout JSON, serialized as a string):

{
"type": "inference.moss-image.sft.txt2img.v1",
"config": {
"prompt": "{\"summary\":{\"meta\":[\"poster\"],\"canvas\":\"single\",\"caption\":\"A minimalist travel poster with a cream background, a blue mountain illustration, and the title EXPLORE at the top.\"},\"items\":[{\"label\":\"text\",\"caption\":\"Large dark-blue uppercase title centered at the top.\",\"text\":\"EXPLORE\",\"bbox\":[-416,-384,-288,384]},{\"label\":\"object\",\"caption\":\"A flat blue mountain illustration in the lower center.\",\"bbox\":[-192,-288,384,288]}],\"bbox_coord_space\":\"abs_centralized\",\"bbox_format\":\"yxyx\"}",
"width": 1024,
"height": 1024
}
}