MOSS-Image-SFT: Generate images with text and objects placed where you say

What it does

MOSS-Image-SFT is a text-to-image model from OpenMOSS. It is a fine-tuned version of MOSS-Image-Base, given extra training (SFT means supervised fine-tuning), using the same design and the same way of running. It accepts three kinds of prompt: plain sentences, lists of tags, or a JSON layout that names each object or piece of text and the box it should sit in. The creators show it producing clearer text grouping and a tidier whiteboard layout than the base model.

How to use it

Input → action → output: write a prompt, or a JSON layout with a caption plus items and their bounding boxes, choose image dimensions, and run the diffusers pipeline. The default settings use 50 steps. You get a generated image, and you can edit the layout's text or boxes and run again to move things around.

Common use cases

• Write a JSON layout listing a title, a chart and two labels with their boxes, and get an image with each one in place.
• Describe a whiteboard full of notes in plain English and get an image where the written text is grouped and readable.
• Use the bundled prompt-enhancer skill to turn a rough idea into a detailed caption, then paste it in to get a richer image.