Previously, in the Lingguang Flash app, users had to wait until app generation was complete before they could see the full page design and visual effects.
In Huamei's professional design scenarios, after designers receive a finished image, they still have to go through layer separation, which typically takes more than 5 minutes before they can continue editing.
One scenario needs to see the design earlier, the other needs to continue editing the design. Both point to the same problem: AI design still needs to move beyond "being able to generate" toward "previewable, editable, and deliverable."
Today, we are officially open-sourcing the Ming-Image-0.1-Design series, which includes two models with 6B parameters each:
-
Ming-Image-0.1-Design (hereafter referred to as Design): generates UI, infographics, posters, and complete visual designs from text requirements;
-
Ming-Image-0.1-Design-Layer (hereafter referred to as Layer): splits creative images or design images into independently editable transparent layers.
We are also open-sourcing Design Skill and PPT Skill, connecting the models' capabilities further to front-end pages and editable presentations.
From generating designs and decomposing layers to entering Coding, PPT, and design tools, what this release hopes to solve is not "generating another image," but how to keep generated results usable.
01 Two 6B models connect design generation with subsequent editing
Design first draws the complete design
UI, Dashboards, and infographics differ from ordinary images. The model needs to handle text, functional modules, visual hierarchy, color schemes, and layout simultaneously; a missing character, a shifted button position, or a confused card relationship can all make the result unusable directly.
Ming-Image-0.1-Design therefore focuses on enhancing four capabilities:
- Support for 8K long structured prompt enhancement, organizing natural language requirements into copy, modules, layout, and visual style;
- Optimized text rendering and layout stability for titles, buttons, cards, and multi-region information;
- End-to-end generation completes the overall design in one pass, unifying color schemes, composition, and asset styles;
- A native RGBA VAE is created to directly generate transparent-background assets such as people, products, icons, and decorations.
Compared with calling general image generation models multiple times through code to separately generate backgrounds, illustrations, and decorations, end-to-end generation has more complete global control and can reduce color clashes and style inconsistencies among assets.
🎬 Video demo (seeoriginal WeChat article ）
Layer turns designs into editable assets
What it does is not ordinary cutout, but restoring the structure behind a design image.
After a design is generated, text will still be modified, people need to be moved, backgrounds may be replaced, and products and decorations also need to be reused separately.
Ming-Image-0.1-Design-Layer can split a design image into 2–9 semantically independent RGBA layers, focusing on scenarios such as advertising posters, PPT, UI/UX, infographics, e-commerce key visuals, and template re-creation.
The model organizes layers according to design roles such as text, cards, subjects, and backgrounds, preserves transparent edges, completes occluded regions, and ensures consistency when all layers are recombined.
🎬 Video demo (seeoriginal WeChat article ）
Why Layer can decompose more accurately
The model uses Type Tokens to tell each layer what design role it plays; Alpha-Aware Layer Optimization focuses on optimizing transparent regions and edges; Composite-Layer Stack Consistency checks through recomposition whether each layer can restore the original image.
In terms of user experience, this means content is separated more clearly, edges are cleaner, and after stacking back together the result is closer to the original design.
Figure 1: Ming-Image-0.1-Design-Layer model architecture. The model combines layer type conditions, Alpha-aware optimization, and recomposition consistency constraints, so that each layer maintains semantic independence and transparent boundaries, and restores the original design as closely as possible after recomposition.
02 6B parameters enter the first tier of design models
Design ranks first among open-source models on the AA UI/UX Design leaderboard
In the UI/UX Design specialized evaluation updated by Artificial Analysis on September 18, 2026, Ming-Image-0.1-Design ranked No. 1 among open-source models with an Elo score of 1082.
In UI/UX scenarios, the model achieved a layout win rate of 67.4%, 67.0% on complex compositions, and 66.7% on text rendering.
These capabilities correspond exactly to the core issues in real pages: whether information can be laid out clearly, whether multiple modules can stay coordinated, whether text can be presented accurately, and whether multiple pages can use the same visual language.
In specific cases:
- For a desktop Dashboard containing an orchard map, picking status, charts, and action buttons, Ming-Image-0.1-Design won 12 / 12 against Nano Banana 2 Lite, FLUX.2 [max], and Krea 2 Medium.
- For a recipe app composed of three consecutive step pages, Ming-Image-0.1-Design won 10 / 10 against Nano Banana 2, Nano Banana Pro, and MAI-Image-2.5-Flash.
This shows that 6B does not mean lowering the requirements for complex design tasks. Through specialized training for text, layout, and multi-element composition, small-parameter models can also enter the competitive range of large models.
Figure 2: Artificial Analysis UI/UX Design open-source model ranking
Figure 3: The orchard dashboard needs to draw an orchard map, picking status, charts, and action buttons at the same time.
Figure 4: A three-step recipe app needs to share the same visual language.
Layer is best on all 12 Crello metrics
For Ming-Image-0.1-Design-Layer, we focus on three questions:
- Whether different content is correctly separated;
- Whether Alpha boundaries are accurate;
- Whether all layers, when stacked back together, are close to the original image.
In the Crello-Test layer decomposition evaluation, Ming-Image-0.1-Design-Layer achieved the best results on both RGB L1 and Alpha soft IoU metrics, as well as under all six settings from MLM-0 to MLM-5, meaning all 12 results ranked first.
Even compared with the non-open-source Qwen-Image-Layered-I2L version fine-tuned on the Crello training set, Ming-Image-0.1-Design-Layer still maintains a comprehensive lead.
Under the same software, hardware, and unified evaluation conditions, the 6B Ming-Image-0.1-Design-Layer outperforms the 20B Qwen official open-source version in quality while reducing single inference time from 795 seconds to 183 seconds, about a 4.3x speedup. After engineering optimization, the current end-to-end service at 1024 resolution typically returns complete results in about 20 seconds.
03 Entering real business, changes go beyond results
Lingguang Flash app: preview designs in advance, final app highly faithful
The Lingguang Flash app lets users generate practically usable mini-apps through natural language.
Previously, during app generation, although the system continuously displayed writing progress, users had to wait until the app was complete before they could see the page layout, color scheme, and overall effect.
After integrating the Ming-Image-0.1-Design series, design previews are moved forward to the initial stage of app generation. Design generates an end-to-end app preview with a complete layout and consistent style based on user requirements, letting users see the app's final visual direction in advance. The Agent then calls Layer to break down the design assets and, combined with the design document, completes code generation, so that the final app is highly consistent with the preview in page structure, visual style, and detail elements, delivering an experience closer to "what you see is what you get."
This change is not just about letting users "see earlier." Design can coordinate layout, color scheme, and asset style from a global perspective, reducing the style fragmentation, template feel, and patchwork feel caused by multiple generations and code assembly, so that AI-generated apps have more complete thematic expression and a more unified design language, closer to a product that has undergone systematic design.
🎬 Video demo (seethe original WeChat article ）
Note: The Flash app integrated with the Ming-Image-0.1-Design series models is currently in internal testing and will be available to everyone soon, bringing a more unified, more controllable, and more efficient app generation experience.
Huamei: from generating an image first and then breaking it down, to obtaining editable layers in about 40 seconds
In professional design, generating an image is usually just the starting point. Copy continues to be adjusted, subjects need to be moved, backgrounds and colors may be replaced, and the design keeps improving as requirements evolve.
The old workflow required generating an image first, then performing cutout, layer separation, and adjustment. It usually took more than 5 minutes to obtain editable layers for the first time.
After integrating Ming-Image-0.1-Design-Layer, Huamei can directly convert images into semantically independent editable layers. Text, subjects, backgrounds, and decorations can be modified, moved, hidden, or replaced separately, and layered results are usually obtained in about 40 seconds.
If only a small amount of text needs to be modified, you can use "text replacement"; when copy needs continuous adjustment, you can use "convert to text"; when you need to move subjects, replace backgrounds, or continue fine editing, you can use "image layering."
Under the current internal testing criteria:
- First-time layer acquisition: from more than 5 minutes down to about 40 seconds, about a 7.5x speedup;
- In the past, completing text replacement with GPT-image2 cost 7 times more than the Ming-Image model.
Design is more than generation. It supports layer editing and copy adjustment, and can export PSD source files, allowing creativity to keep improving.
🎬 Video demo (seethe original WeChat article ）
Note: Huamei is an intelligent design application within Ant Group for designers and business teams, supporting capabilities such as AI image generation, copy adjustment, image layering, layer editing, and PSD export.
From model capabilities to front-end and office delivery
This time, Design Skill and PPT Skill are also open-sourced.
Design Skill Supports Text-to-Page and Image-to-Page（Visual Coding）connecting Design, Layer, and Coding into a complete pipeline that directly generates runnable front-end pages that can be further modified. Traditional AI Coding often generates code first and then opens the page to see the effect; if the style or layout is wrong, you have to go back and modify the code or replace assets, and when changes are large, you may even need to regenerate. Text-to-Page moves this process forward to "design first, then write code": the user enters a requirement, and in about 15 secondsthey can first see a complete page plan; after confirming the style and layout, assets are extracted and handed to Coding to build the page. This way, the time to first see a result can drop from about 5 minutes to 15 secondschoosing the right direction before starting Coding and reducing ineffective generation and subsequent rework.
Visual Coding is "restoring a page directly from a reference image": after inputting a design draft or screenshot, the Layer model first understands the page structure and visual relationships, correctly breaks down assets such as backgrounds and illustrations, and then enters the Coding stage; after the page is generated, it also usesscreenshots to perform visual verification against the original image and continue correctingCompared with traditional "look at the image and write code," Design Skill turnsvisual understanding, asset decomposition, code generation, and visual verificationinto a closed loop, improving page restoration accuracy while reducing the cost of repeatedly adjusting code and replacing assets.
🎬 Video demo (seethe original WeChat article ）
PPT Skill can restore the text, color blocks, and layout relationships in an image into editable elements in PowerPoint. Visual assets needed for replication, such as icons and illustrations, can be directly separated and extracted from the source image without additional repeated generation, preserving the original visual details while allowing users to continue modifying copy, fonts, color schemes, images, and page layout.
Used to directly restore a design image into an editable PowerPoint page. It recognizes the text, color blocks, images, and layout relationships in the image and converts them into individually editable elements in PPT; visual assets such as icons and illustrations can be directly extracted from the original image for reuse without regeneration. After conversion, users can continue to modify copy, fonts, color schemes, images, and page layout, turning a "view-only" image into a "further editable" PPT page while preserving the original visual effect as much as possible.
🎬 Video demo (seethe original WeChat article ）
The four capabilities can be used independently or in combination:
Text requirement or reference image → Design generates a complete visual → Layer separates independent assets → continue editing and reuse → generate a front-end page or editable PPT
For external developers, the 6B parameter size lowers the barrier to deployment and secondary development; the simultaneous open-sourcing of model weights and Skills provides a set of foundational capabilities from visual generation to application delivery. Developers can build design tools, merchant marketing products, office applications, and front-end generation services based on it.
Current limitations and future plans
The Ming-Image-0.1-Design series still has clear capability boundaries.
Design is currently most stable with typography and text; however, when encountering complex hand movements, continuous operation steps, or the need for very fine shadows and reflections, the results are not yet stable enough.
In the Layer model, there is no single answer for layering. Some people only need three layers: background, subject, and text, while others want each piece of text and each icon separated individually. Heavy overlapping light effects, special transparent materials, and complex occlusion may also cause edge residue or missing layers.
Next, we will continue to improve text and complex layout stability, expand layer semantics and controllable layering capabilities, optimize transparency effects, occlusion completion, and inference efficiency, and integrate the models into more design, office, and application generation scenarios.
What we care about is not just what the model can generate, but whether the generated results can truly move to the next step: continue to be modified, continue to be used, and continue to be delivered.
04 Try it now
Online experience
To make it easy for developers to try quickly, Ming-Image-0.1-Design (https://openrouter.ai/inclusionai/ming-image-0.1-design and Ming-Image-0.1-Design-Layer (https://openrouter.ai/inclusionai/ming-image-0.1-design-layer are available for free API calls on OpenRouter for two weeks.
🤗 Hugging Face
https://huggingface.co/inclusionAI/Ming-Image-0.1-Design
https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer
🤖 ModelScope https://www.modelscope.cn/models/inclusionAI/Ming-Image-0.1-Design
https://www.modelscope.cn/models/inclusionAI/Ming-Image-0.1-Design-Layer
Design Skill https://github.com/inclusionAI/ling-cookbook/tree/main/resources/recommended-skills/ling-ui-design
If you are creating UI, infographics, posters, presentations, or design assets, or if you already have a design image you want to keep editing, you are welcome to try the Ming-Image-0.1-Design series.
You are also welcome to share with us the cases from real projects that are hardest to generate, hardest to modify, or hardest to layer.
Let AI not only generate an image, but also understand design and continue to deliver the design.
Model releaseMultimodalPrevious postLing-3.1-flash: A model upgrade for real-world tasksNext postImproving the intelligence efficiency of large models starts with shortening the wait for a single data validation