openmoss/moss-image is the official repository for MOSS-Image, which generates images from text. Built on a 5B-parameter diffusion transformer (DiT), it emphasizes text rendering for lettering inside images and control over the placement of text and objects. The code and model weights are released under the Apache-2.0 license.
The README introduces the MOSS-Image-Base and MOSS-Image-SFT weights. You can specify the positions of in-image text and objects with JSON layouts, and it also supports long prompts and absolute layout input based on pixel coordinates. The repository description states that prompts of up to 4K tokens are accepted. It provides Python and Gradio usage plus a ComfyUI plugin. The weights are distributed on Hugging Face, and the pipeline class MOSSImagePipeline is registered in diffusers. The emphasis the repository puts forward is image generation with public weights and layout control, rather than the training method.
That said, the repository contains no training code, and its contents center on an inference wrapper, a Gradio app, a ComfyUI plugin, and website assets. The README highlights efficient fine-tuning, but whether fine-tuning can be performed with this repository alone has not been confirmed. No figures backing text accuracy or image quality are given within the repository either. There are 10 recorded commits, and the provided information lists no latest release. So the published features can be checked, but the basis for judging performance and the training path is limited.