It is a 5B text-to-image model for readable English and Chinese lettering, prompts of up to 4,096 tokens, and JSON layout. Anime and photographic styles are both in range. Apache-2.0 covers the weights and the inference code. You still answer for trademarks, likenesses, and copyrighted characters in a prompt.
Try MOSS Image online
The form calls MOSS Image on the official Hugging Face Space. There is no account on this page. Free runs use Hugging Face ZeroGPU, a shared pool, so a queue is normal. A 50-step picture near one million pixels often takes half a minute to a couple of minutes. Hugging Face caps free GPU time per visitor per day, so sketch with fewer steps before a full run.
Leave More settings closed for the official starting point: 50 steps, CFG 3, shift 5, and a random seed. Shape buttons stay near one million pixels, with both sides multiples of 16. Put words that must be drawn inside quotation marks, and put each line of a sign or poster on its own line. The button runs the SFT checkpoint. A community Space can make an 8-step draft. OpenMOSS has not checked that shortcut. If the form cannot reach the Space, use the full-screen demo under the button.
A short guide to MOSS Image
Use the form when you only need a picture. Move to ComfyUI when you want a GPU you control. Clone the official repository when you want the pipeline the authors ship.
In the browser
Write the prompt, pick a shape, and press Generate. The picture shows in the frame and can be downloaded. If the daily free cap is spent, sign in at Hugging Face or wait for the reset. This page does not store the error or the prompt. The authors' panel is the same Space opened full screen.
On a 12 GB GPU with ComfyUI
A 12 GB card is the practical floor for the smaller files. The three files below are the SFT set from Comfy-Org. Together they are about 13 GB. The Comfy-Org repo also hosts larger BF16 copies, which need more than 12 GB.
| File | Folder | Size |
|---|---|---|
| moss_image_sft_int8_convrot.safetensors | diffusion_models | 5.4 GB |
| qwen3.5_9b_w4a8_gptq.safetensors | text_encoders | 7.3 GB |
| flux2-vae_small.safetensors | vae | 0.25 GB |
In ComfyUI Manager, search for the project plugin if you want the authors' loader. Nightly builds also include native support from the release day. Load the diffusion file, the text encoder, and the small FLUX.2 VAE. For a first picture use Euler, 50 steps, CFG 3, and shift 5. Skip the negative branch at layer 22. A 1024 by 1024 latent is a sound test.
From the official repository
Clone the OpenMOSS repository and follow its README for the Python environment, the requirements file, and the local Gradio command. This page does not repeat that script. Use the SFT weights when you want a finished MOSS Image picture. Base is the MOSS Image checkpoint kept for LoRA training, because it has not had the curated fine-tune. The same width, height, step count, CFG, shift, and layer skip apply in the script, in ComfyUI, and in the form. Both checkpoints are on Hugging Face for SFT and Hugging Face for Base.
Settings worth keeping
Keep shift at 5 in ComfyUI and in the local script. CFG can sit between 3 and 6. This form defaults to 3. Layer 22 is the negative-branch skip. Leave it on for a first test. Sides must be multiples of 16, which the shape buttons already are. A useful size is about 1 million pixels. The model can run from about 0.25 million pixels up to about 4 million.
Layout, when you need it
A prompt can be JSON instead of a sentence. Describe the scene, then give a box for each line and each object. Paste that JSON into the same prompt box, and match its width and height to the selected shape. The project page documents the coordinate format. Layout does not need a second model.
Common questions
Can MOSS Image edit a picture?
No. This release is text-to-image only. Editing is not in the weights you can download today.
Can I use the pictures commercially?
The code and the weights are Apache-2.0, which allows commercial use of the model. You still have to clear trademarks, real people, and copyrighted characters yourself. This page is not legal advice.
Will a Chinese prompt work?
Yes. English is the larger share of the training data, and Chinese works both as the prompt and as text drawn in the picture. Official examples use both.
Why does Generate wait?
The free GPUs are shared. Busy hours add waiting time, because MOSS Image is sharing those GPUs. Sampling still runs after the queue reaches you. Fewer steps return sooner. The picture still comes from the official Space.