Ming-Image-0.1-Design is a 6B open text-to-image model from Ant Group's inclusionAI lab, built for UI, infographics, posters and other text-rich designs. It's released under the MIT License, and it can output RGBA images with transparent backgrounds.
What you need
- One CUDA GPU with 80 GiB of VRAM, running BF16. That's the only setup Ant validated, so plan on a cloud GPU rather than a laptop.
- Python and git.
Install
git clone https://github.com/inclusionAI/Ming-Image
cd Ming-Image
pip install -r requirements.txt
Run
python infer.py --model inclusionAI/Ming-Image-0.1-Design --task text-to-image --prompt assets/t2i_four_seasons_cabin_prompt.json --resolution 2048 --output-dir outputs/t2i
Swap in your own prompt file once the example works.
Recommended settings
| Setting | Value |
|---|---|
| Resolution | 2048 x 2048 (or 1024 x 1024 for faster generation) |
| Sampling steps | 12 |
| CFG scale | 1.0 |
| Precision | BF16 |
Readable text: the rule
In structured JSON prompts, put every string you want printed in quotes, exactly as it should appear, and only once. Don't repeat the same words in other description fields.
Transparent background: the phrases
Start your prompt with exactly one of these:
RGBA, 4-channel, transparent backgroundisolated subject, alpha matte, no backgroundcutout PNG, alpha=0 outside the objecttransparent canvas, not white, not checkerboard带透明通道，4通道RGBA图像(Chinese)
Split a design into layers
The companion model Ming-Image-0.1-Design-Layer (also 6B, MIT) takes a flattened design plus a layer plan and splits it into 2 to 9 editable transparent RGBA layers. It's the closest thing to getting a working design file back.
Good to know
- It's text-to-image only. It doesn't take reference images.
- It's the #1 open-weights model on the Artificial Analysis UI/UX Design leaderboard (#17 of 81 overall), but only #45 on the general text-to-image board.
- Ant hasn't published results for text in languages other than English, so test your own before you rely on it.