Ming-Image-0.1-Design is a 6B text-to-image model built for visual design work: UI mockups, infographics, posters, slides and anything else where the text has to come out right. It renders legible type as part of the composition instead of the usual generative-model scribble, and it can output RGBA with a transparent background.
Originally released by inclusionAI on Hugging Face under the MIT license. All credit for the model goes to the inclusionAI team. Civitai is hosting a mirror so creators can run it on-site without a local install - please head to the original repository for weights, updates and to follow the project directly.
Built by
- inclusionAI - the Ming-Image series and the Ling-mini-2.0 text encoder it runs on
- The Ming-Image repository - reference inference code, prompt rewriting and the layer-decomposition demo
- Comfy-Org - the repackaged single-file weights this mirror is built from
Text-rich design, not just pictures
The model was trained for complete visual compositions rather than single subjects. Prompts that describe a layout - headings, body copy, callouts, spacing - are what it is built to answer, and it places readable text inside that layout. It sits near the top of the Artificial Analysis UI/UX design leaderboard.
Transparent backgrounds
Ming-Image-0.1-Design generates RGBA output with genuine alpha, so a logo, sticker or UI element comes out ready to composite. Upstream recommends prepending one of the RGBA trigger phrases documented in the transparent-background tip to switch it on.
Recommended settings
- Resolution: 2048 x 2048 for best quality, 1024 x 1024 for speed. Output is square at the selected bucket.
- Steps: 12.
- CFG scale: 1.0. Negative prompts do nothing at CFG 1.
- Precision: BF16.
Upstream also suggests running prompts through a rewriter first - Ling-3.0-flash-VL or a comparable VLM - since the model responds well to long, structured design briefs.
The Layer model
The series has a second 6B model, Ming-Image-0.1-Design-Layer, which takes a flattened design and splits it back into separate transparent layers you can edit. It is a different checkpoint with its own task and is not part of this model page.
Links
- Source: inclusionAI/Ming-Image-0.1-Design on Hugging Face
- Code: github.com/inclusionAI/Ming-Image
- ModelScope: inclusionAI/Ming-Image-0.1-Design
- Demo: Hugging Face Space
- Agent skills: Ling UI Design and Image to Editable PPT
Description
FAQ
Comments (16)
34 GB of (TE)? DAMN, THAT'S BIG
Int8 is 18gb, int4 is 12gb
@Fatbuns no chance for 8gigs.
Int8 18gb seems whack, but tbf, it runs fine on 12gb VRAM. int4 12gb......yeh, I probably wouldn't bother at that point.
It also likes verbose prompting style; so you probably need a side/their side LLM to edit you original prompts, too.
Z-image turbo v1.5
XDD
the Loras of Z-Image work with Ming Image
hello, where to find a workflow please ? images are without metadata. thanks !
https://github.com/Comfy-Org/ComfyUI/pull/16482 examples here, but you can generally just make it yourself, if you have basic knowledge about comfy
You can try to use mine, should be compatible : https://civitai.red/models/2961450/ming-image-01-design-text-to-image-workflow-comfyui
@noctaluna Hi Thank you. Have a good week end
@Fatbuns Hello thanks for your link. Have a good Week end.
Z-Image LORAs can be used with Ming Image
Does this model work with any version of Qwen TE???
Or it only works on Ling-TE? :((
That Ling-TE@12gb is too big for my old hardware :(
I hope they come out with TE@5-6gb, please!!!!
It works with the TE used for Z-image, I have generated an image using qwen_3_4b.safetensors but with important cost you can see here an example using
https://civitai.red/posts/31384978 qwen_3_4b.safetensors, while using the MIng TE I got https://civitai.red/images/144144663
1K is much better than 2K results. it is kind of wrong