inference.moss-image.sft.txt2img.v1
The inference.moss-image.sft.txt2img.v1 job generates an image from a text
prompt with MOSS-Image, OpenMOSS’s open-weight text-to-image model. It is
strong at rendering long English and Chinese text and accepts JSON layouts
that place text and objects at exact positions. Output is 1024x1024 by
default.
{ "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!\"" }}Prompting
Section titled “Prompting”The prompt is used exactly as written. It can be natural language (English
works best, Chinese and other languages are supported), a tag sequence, or a
JSON layout. Put the exact text to render in quotes and use line breaks
(\n) where the text should break. The model reads at most 4096 tokens
(about 16,000 characters of English, fewer for Chinese); anything beyond that
is cut off.
negative_prompt (empty by default) describes what to avoid, e.g.
worst quality, bad anatomy, watermark, jpeg artifact.
guidance_scale (1-10, default 3) sets how strongly the prompt is followed:
3 suits long prompts, short prompts can benefit from up to 6.
Layouts
Section titled “Layouts”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 1024x1024 image both axes run
from -512 to 512. Send the same width and height the layout was drawn
for.
width and height can each be any integer from 512 to 2048 (default
1024). The model works best at 1-4 megapixels.
The same seed reproduces the same image.
Latency
Section titled “Latency”Typical latency on a warm worker is about 4.5 s at 1024x1024 and 25 s at 2048x2048.
Schema
Section titled “Schema”{ "type": "object", "required": [ "type", "config" ], "additionalProperties": false, "properties": { "type": { "type": "string", "enum": [ "inference.moss-image.sft.txt2img.v1" ] }, "config": { "type": "object", "required": [ "prompt" ], "additionalProperties": false, "properties": { "prompt": { "type": "string", "minLength": 1, "maxLength": 16384, "examples": [ "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!\"" ], "description": "The prompt to generate an image from, used verbatim: natural language, a tag sequence or a JSON layout. Text beyond the model's 4096-token limit (about 16,000 characters of English, fewer for Chinese) is cut off." }, "negative_prompt": { "type": "string", "maxLength": 16384, "default": "", "examples": [ "worst quality, bad anatomy, watermark, jpeg artifact" ], "description": "What to steer away from. Empty by default; also cut off at 4096 tokens." }, "width": { "type": "integer", "minimum": 512, "maximum": 2048, "default": 1024, "description": "Output image width in pixels, 512 to 2048." }, "height": { "type": "integer", "minimum": 512, "maximum": 2048, "default": 1024, "description": "Output image height in pixels, 512 to 2048." }, "guidance_scale": { "type": "number", "minimum": 1, "maximum": 10, "default": 3, "description": "Guidance scale for classifier-free guidance. 3 suits long prompts; short prompts can benefit from up to 6." }, "seed": { "type": "integer", "minimum": 0, "maximum": 2147483647, "description": "Random seed for reproducible results." } } } }}