Image creation
MOSS-Image
An open-source 5B text-to-image model with sharp Chinese and English text and layout control by bounding box.
Product insights
MOSS-Image at a glance
Value proposition
A free, self-hostable image model that renders legible text and lets you place every element exactly where you want it.
Problem solved
Most open image models garble text, especially Chinese, and give no direct way to control where things appear on the canvas.
Audience
Designers making posters and infographics, ComfyUI users, and developers or researchers who want an open base model to fine-tune with LoRA.
Pricing and access
Free and open source. Weights are released under Apache-2.0; you pay only for your own GPU or hosting.
Market
- Open-sourced on GitHub (OpenMOSS/MOSS-Image) and Hugging Face on October 16, 2026
- Free online demo on Hugging Face Spaces
Tech stack
- 5B diffusion transformer with a Qwen3.5 text encoder and FLUX.2 VAE
- Layout JSON prompts with per-item bounding boxes and text
- Diffusers pipeline, Gradio app, ComfyUI plugin, and vLLM-Omni serving
01Why it is here
What is MOSS-Image?
MOSS-Image is an open-source 5B text-to-image model from OpenMOSS that renders Chinese and English text well and lets you control layout with bounding-box JSON prompts.
02Workflow
How to use MOSS-Image
- 01
Try the online demo
Open the MOSS-Image Space on Hugging Face and type a prompt to generate an image without installing anything.
- 02
Run it locally
Clone the OpenMOSS/MOSS-Image repo and install the requirements, then use the Python pipeline or launch the Gradio app. ComfyUI users can install the plugin from ComfyUI Manager.
- 03
Control the layout
For posters or infographics, write a layout JSON that gives each element a caption, a bounding box, and the exact text to draw.
03Capabilities
Key features of MOSS-Image
- Clear Chinese and English text rendering
- Layout control with bounding-box JSON prompts
- Long prompts up to 4K tokens
- Base checkpoint for LoRA fine-tuning and SFT checkpoint for direct use
- Anime and photorealistic styles
- Native ComfyUI plugin
04Use cases
What is MOSS-Image best for?
- Posters, infographics, and covers with readable text
- Designs where elements must sit in exact positions
- Fine-tuning a small open image model on your own style
05Questions
FAQ about MOSS-Image
What is MOSS-Image?
MOSS-Image is an open-source text-to-image model from the OpenMOSS team. It is built on a 5B diffusion transformer and is known for clear Chinese and English text and layout control.
Is MOSS-Image free?
Yes. The weights are open under Apache-2.0, and you can try it free on Hugging Face Spaces or run it on your own GPU.
What is the difference between MOSS-Image-Base and MOSS-Image-SFT?
Base is the raw pretrained model, meant as a starting point for LoRA and other fine-tuning. SFT is tuned on curated data and gives cleaner text grouping and layouts out of the box.
Can I use MOSS-Image in ComfyUI?
Yes. A native ComfyUI plugin is available through ComfyUI Manager, with an example workflow included.
Official website snapshot
openmoss.ai
MOSS-Image is an open-source text-to-image model family from the OpenMOSS team, released on October 16, 2026. It pairs a 5B-parameter diffusion transformer with a Qwen3.5 text encoder and a FLUX.2 VAE. Besides plain prompts of up to 4K tokens, it accepts a layout JSON that places each object and each line of text in its own bounding box, which makes it well suited to posters, infographics, and other designs with readable Chinese or English text. It comes in two checkpoints, Base for fine-tuning and SFT for direct use, and runs in Python, a Gradio app, ComfyUI, or a free Hugging Face demo.
