On October 16, 2026, the OpenMOSS team released MOSS-Image — a 5-billion-parameter open-source text-to-image diffusion model. In the week and a half since, it has drawn attention for one reason above all others: its project emphasizes Chinese and English text rendered directly inside generated images, alongside structured layout control.
But MOSS-Image is more than its text capability. It ships with native ComfyUI support, a JSON layout system for structured composition, two checkpoints for different use cases, and an Apache 2.0 license that permits commercial use. This article explains what it is, who built it, what makes it different, and how to start using it today.
Who made MOSS-Image
MOSS-Image comes from OpenMOSS, the same research group behind the MOSS large language model series. The project is hosted on GitHub under the OpenMOSS organization and the model weights are published on Hugging Face under both the OpenMOSS-Team and Comfy-Org accounts.
The model is not affiliated with mossimage.run — this site is an independent online studio built by Moxion Labs LLC that hosts the SFT checkpoint through fal.ai's inference API. Think of it as the relationship between Stable Diffusion and the many websites that let you generate images with it. The model is open; the studio is one way to use it.
What makes MOSS-Image different
Most AI image generators share a common weakness: text. Ask them to put a price on a menu, a title on a poster, or a label on a product, and the letters often come back misspelled, garbled, or replaced with something that looks like text from a distance but is not. Chinese characters fare even worse — most models produce shapes that resemble characters but are not real ones.
MOSS-Image was trained with text rendering as an explicit objective. The official project description highlights "text rendering and layout control," and the published examples include English text, Chinese text, and mixed-language designs. The project treats text rendering as a core capability rather than a secondary demo.
The other capability that sets it apart is JSON layout control. Instead of describing an image with spatial adjectives ("put the title at the top"), you give each text or object element a bounding box measured in pixels from the canvas center. The model reads that structure as layout guidance. It is experimental, and the final output can still vary, but it is a fundamentally different approach to composition than free-form prompting.
A third practical advantage: native ComfyUI support from day one. MOSS-Image support was merged into ComfyUI on the same day as the model release. There is no waiting for community nodes or reverse-engineering the pipeline — the workflow graphs, loader nodes, and model components are documented and available.
The two checkpoints: Base and SFT
MOSS-Image ships two versions of the model, and choosing between them matters:
| Base | SFT | |
|---|---|---|
| What it is | The foundation checkpoint | Supervised fine-tuned from Base |
| Best for | Evaluation, LoRA training, custom post-training | Direct image generation |
| Text rendering | Supported but needs careful settings | Better text grouping in published comparisons |
| JSON layout | Supported | Supported |
| License | Apache 2.0 | Apache 2.0 |
SFT is the right choice for most people. It is the checkpoint you use when you want to generate an image now. Base is for researchers, LoRA trainers, and anyone who wants to build their own post-training process on top of the foundation model.
The hosted studio at mossimage.run uses the SFT endpoint through fal.ai. If you want to run the model locally, both checkpoints are available on Hugging Face with matching ComfyUI workflows.
The relationship between Hugging Face and mossimage.run
If you search for "moss image" on Hugging Face, you will find the official model repositories under OpenMOSS-Team and Comfy-Org. These are the source of truth for model weights, code, and documentation. You can download the checkpoints, run them locally in ComfyUI, and use them in your own projects under the Apache 2.0 license.
mossimage.run is a separate, independent online studio. It does not host the model itself — it sends generation requests to fal.ai's hosted MOSS-Image SFT endpoint. The studio provides a browser-based interface, prompt examples, a gallery of real outputs, and guides for getting started. It is one of several ways to use the model, alongside running it locally in ComfyUI or calling the fal.ai API directly.
This is an important distinction because it affects cost, privacy, and control. The online studio charges for image credits (with three free starter images). Running locally requires compatible hardware and carries its own hardware, electricity, and maintenance costs. The fal.ai API charges per megapixel. Each option serves a different use case.
What MOSS-Image can do — and what it cannot
What it does well:
- Readable English text at short lengths — headlines, labels, menu items, prices
- Chinese character rendering — not perfect, but positioned specifically for this task
- Bilingual designs with separate Chinese and English text elements
- Structured layouts through JSON bounding boxes
- Repeatable testing with recorded seeds and settings
- Integration with ComfyUI workflows
What it cannot do:
- Guarantee correct text on every generation — proofreading is still required
- Replace a design tool — JSON layout guides composition, it does not give pixel-perfect typography
- Generate long paragraphs of text — short strings work, dense blocks do not
- Edit existing images — it is a text-to-image generator, not an editor
The model card and user guide are refreshingly honest about these limitations. "Layout prompts are experimental" appears in the official documentation. The comparison with FLUX on mossimage.run shows both successes and failures, with no cherry-picking. This transparency is worth noting because it sets realistic expectations.
How to try MOSS-Image today
You have three paths, depending on what you need:
1. Browser-based studio (fastest). Go to mossimage.run, verify your email, and generate three images for free. No GPU, no setup, no credit card. This is the best way to test prompts and see what the model can do before committing to a local install.
2. Local ComfyUI (most control). Download the model components from Hugging Face, import the native ComfyUI workflow, and generate on your own GPU. A full step-by-step guide is available. This path gives you direct control over settings and local files, while requiring compatible hardware, electricity, setup, and maintenance.
3. fal.ai API (for developers). Call the hosted endpoint directly if you are building an application. Pricing is $0.05 per megapixel, which works out to about $0.0524 for a 1024 × 1024 image. API documentation is available on fal.ai.
More guides: ComfyUI setup guide · Local install guide · Base vs SFT comparison
Official sources (checked October 26, 2026): MOSS-Image GitHub · MOSS-Image SFT on Hugging Face · OpenMOSS user guide · ComfyUI setup guide
Three free images. No setup required. [Open the studio →](/#studio)