Generate images with readable text and a layout you control

You describe a scene and include the exact words you want in it. It returns an image with that text lettered in: a poster title, a café menu board, a sticky note in handwriting on a fridge. Prompts can be long and detailed, so you can spell out every line of a flyer instead of hoping the model fills in the gaps. English works best, and Chinese text works too.

For tighter control, you can describe the layout as a list of elements, each with its own box on the canvas: the title across the top, an illustration in the middle, a price tag in the corner. The model puts the text and objects in those spots. The layout is a short block of structured text you paste into the prompt, and the project's user guide has a ready poster example to copy.

This is the fine-tuned version of the MOSS-Image family from the OpenMOSS team, trained further on curated images so text groups more cleanly than in the base model. It is a 5 billion parameter model under the Apache 2.0 license, so you can use it and its images in commercial work.

The demo Space opens with a strip of example images. Click one to load its prompt and settings, then type sft into the Model field so you are running this version. Now write your own: "A chalkboard menu outside a café that reads: Flat white 3.50, Croissant 2.80, Open until 6." Read every word on the result, then ask for the same board with a short Chinese line underneath. To run it yourself, it works with Diffusers on an NVIDIA GPU and has a ComfyUI plugin.

Long passages and line breaks do not always land where you expect. Putting the line breaks directly in your prompt helps. The demo gives visitors who are not signed in only a small shared GPU allowance, so sign in to Hugging Face if it keeps queueing.

• Mock up a poster or event banner with the real title and date already in it.
• Make menu boards, shop signs and price tags for a storyboard or pitch deck.
• Create bilingual English and Chinese visuals for packaging or social posts.
• Generate a product shot with the label text you choose.

Try the demo

View model page

Read this issue on neuralexpedition.com →