What Is MOSS-Image? The Open Text-to-Image Model Explained

Independent site. Verdapix is not affiliated with, endorsed by, or sponsored by the OpenMOSS team or the authors of MOSS-Image. Facts below are summarized from the project's public pages as of October 20, 2026. Where something could not be confirmed, we say so.

MOSS-Image in one paragraph

MOSS-Image is a family of open-weight text-to-image models published by the OpenMOSS team. According to the official project page and GitHub repository, the first checkpoints — MOSS-Image-Base and MOSS-Image-SFT — were released on October 16, 2026. The model is described as a compact 5B-parameter diffusion transformer whose main strengths are rendering readable text inside images, following long prompts (up to about 4K tokens), and layout control through structured JSON prompts.

Key features

Accurate text in images

The headline feature is text rendering. You can describe a sign, poster headline or label and put the exact words in quotes, and the model tries to reproduce them legibly. The official materials highlight both English and Chinese text. Like every image model, it still makes mistakes — always proofread the result.

Long prompts and layout control

MOSS-Image accepts long, detailed prompts. The project also documents a JSON layout format in which you list objects and text blocks with bounding boxes to control where each element appears. Our generator currently uses plain-language prompts; layout prompts are on our roadmap.

Two checkpoints: Base and SFT

• MOSS-Image-Base — the base text-to-image model, intended as a starting point for fine-tuning and LoRAs.
• MOSS-Image-SFT — fine-tuned from Base on curated data. The model card says it produces cleaner text grouping and layouts.

Based on the inference provider's model description, this website runs the SFT variant.

Open license

The project states that the model weights are released under the Apache-2.0 license. Check the official model cards for the exact terms that apply to each file.

How to use MOSS-Image on this website

1. Open the generator and pick a template: Poster, YouTube Thumbnail or Free generation.
2. Describe your image. Put any text that should appear in the image inside quotes, for example: a café sign that reads "CORNER CAFE".
3. Choose a size and how many images you want, click Generate, then download the result.

New accounts get 3 free images. Each image costs 1 credit.

Prompt tips

• Quote the exact text you want rendered, and keep it short — headlines of a few words work best.
• Describe placement: "large headline at the top", "small caption at the bottom right".
• Name the style: "minimalist Swiss typography", "photorealistic", "flat vector illustration", "anime style".
• Pick the right shape: portrait for posters, 16:9 for thumbnails and banners.

Run MOSS-Image yourself

Because the weights are open, you can run MOSS-Image on your own GPU. See our ComfyUI guide for an overview of the local setup.

FAQ

Who made MOSS-Image?

The OpenMOSS team, according to the official project page (openmoss.ai/MOSS-Image) and the OpenMOSS GitHub organization. This website is an independent service and is not run by that team.

Is MOSS-Image free?

The model weights are openly available. Running it in the cloud costs compute, so this website gives every new account 3 free images and will offer paid credit packs later.

Can MOSS-Image edit existing photos?

Not that we could confirm. The official materials describe text-to-image generation only, and this website only supports text-to-image.

Can I use the images commercially?

The weights are published under Apache-2.0. Your use must also respect our Terms of Service, our Acceptable Use Policy, and other people's rights, such as trademarks and likenesses.

How accurate is the text rendering?

Usually good for short English and Chinese phrases, less reliable for long paragraphs or unusual words. We have not run a formal benchmark — treat claims about accuracy as approximate.