OpenMOSS/MOSS-Image

What it claims

MOSS-Image is a 5B-backbone text-to-image family released on 16 October 2026 as MOSS-Image-Base and MOSS-Image-SFT, with code and weights public. A technical report followed on arXiv on 19 October 2026 (eprint 2610.23374, credited to MOSI. AI). The README claims layout control, accurate in-image text rendering (its example renders handwritten Chinese), prompts up to 4K tokens, anime and photorealistic styles, and training on "alternative GPU platforms". It does not name those platforms. In the README's comparison, SFT gives clearer text grouping and a more organised whiteboard layout than Base.

How to use it

Install from source:

git clone https://github.com/OpenMOSS/MOSS-Image.git
cd MOSS-Image
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

The Diffusers pipeline loads a custom pipeline with trust_remote_code=True. It keeps the VAE in float32 and the rest in bfloat16. The README example uses 1456×1456, 50 steps, guidance_scale=3.0 and uncond_skip_layers=[22]:

pipe = DiffusionPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-Image-Base",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)

The Gradio app:

pip install -e '.[ui]'
python gradio_app.py --model OpenMOSS-Team/MOSS-Image-Base

Gradio has a Memory mode setting. On low GPU memory, choose Sequential or reduce image size. Compile DiT is off by default and does not work in Sequential mode.

Layout control takes a JSON prompt with three parts:

• a summary holding meta, canvas and caption;
• items, each with label, caption, bbox and an optional text to render;
• "bbox_format": "yxyx", meaning [y_min, x_min, y_max, x_max].

With "bbox_coord_space": "abs_centralized", coordinates are pixels relative to the canvas centre. On 1024×1024 each axis runs from −512 to +512, and y increases downward. To convert ordinary pixel coordinates, subtract W/2 from x and H/2 from y. The output width and height must match the coordinates. From Python, pass json.dumps(layout, ensure_ascii=False) as the prompt.

The ComfyUI plugin installs through ComfyUI Manager by searching "MOSS-Image". It is not yet in the official ComfyUI repository.

The Prompt Enhancer is an agent skill invoked as $moss-image-prompt-enhancer. It targets GPT-6-family agents with local file and Python tools. Its script turns pixel-coordinate drafts into compact centred JSON and does not call an image service.

Limits

The page reports no benchmark numbers, inference speed or GPU memory figures. Apache License 2.0 covers "the standalone pipeline code". The page does not state a licence for the model weights, so check it on the model card before commercial use. Loading requires trust_remote_code=True, which runs code from the model repository. The page does not show the full Memory mode list, and it says nothing of the SFT model's training data or of how Base differs from SFT beyond the one example.