Ming-Image targets readable text and editable design assets

Ant Group’s inclusionAI team has released Ming-Image-0.1-Design, a 6-billion-parameter text-to-image model built for interfaces, dashboards, infographics, posters and other text-heavy graphics. It ranks first among open-weight models on the UI/UX Design board from Artificial Analysis, placing 17th among 81 models overall in that category.

The ranking reflects a narrow strength. Ming-Image-0.1-Design ranks 45th on the broader Text to Image leaderboard, with a displayed Elo score of 995, compared with 1,082 for UI/UX Design. Artificial Analysis calculates Elo from head-to-head preference votes, so each category’s prompts determine what the score measures. The results position Ming-Image as a specialist for structured visual design rather than a leading general image generator.

Four pieces in the release

The release contains two downloadable 6B checkpoints and two workflow tools. Open-weight access allows teams to inspect, host and adapt the parameters within the licence terms. The project uses the MIT licence, which permits commercial use, modification and redistribution while requiring preservation of its copyright and licence notices; third-party dependencies retain their own terms.

Higher scores where text matters

Artificial Analysis also evaluates models across capability slices that differ from its use-case boards. Ming-Image is the highest-ranked open-weight model for Layout, at 19th among 67 models overall, and Text Rendering, at 23rd among 65. Layout tests elements such as arrows, visual hierarchy and multi-panel composition. Text Rendering covers long copy, small type, symbols and stylised lettering.

The displayed UI/UX score sits 87 points above the general score, with minor discrepancies possible when the leaderboard rounds underlying values. Productivity and Knowledge Work, which includes diagrams, charts, infographics and slides, scores 48 points above the general result. These category gains explain why structured design briefs suit the model better than broad image prompts. Rankings remain dynamic as Artificial Analysis collects additional votes.

Compute sets the deployment boundary

The model card recommends outputs at 2,048 × 2,048 pixels, with 1,024 × 1,024 available for lower-cost runs, using 12 sampling steps and a classifier-free guidance scale of 1.0. Sampling steps specify the number of denoising passes. Classifier-free guidance normally amplifies prompt conditioning; in conventional implementations, a scale of 1.0 adds no extra amplification. The published quality results apply to this tested recipe, so alternative settings require separate evaluation.

The validated single-GPU configuration requires an H100, H200 or A100 with 80GB of memory. That requirement places local inference on datacenter hardware and makes compute cost a deployment consideration even though the licence carries no usage fee. No first-party or major third-party hosted endpoint was available at publication.

The public Ming-Image repository provides the reference implementation. Its command-line example installs the project requirements and runs a 2,048-pixel text-to-image job:

git clone https://github.com/inclusionAI/Ming-Image
cd Ming-Image
pip install -r requirements.txt
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

Transparent-background generation requires one of the repository’s supported RGBA phrases at the beginning of the prompt. The supplied prompt templates contain the exact wording. For ComfyUI, Kijai has published repackaged BF16, int8_convrot and w4a8 variants of the transformer and text encoder. The quantised variants reduce memory use by storing some values at lower precision, with compatibility and output quality dependent on the selected build. Proposed core support, including RGBA output, appears in ComfyUI PR #16482.

Layers make the output editable

Design-Layer addresses a practical limitation of generated graphics: a flattened PNG merges every element into one pixel grid. Moving a headline, replacing a logo or adjusting a background then requires image editing or another generation pass, either of which can disturb unrelated elements.

The layer model accepts a finished image and a plan specifying the desired layer count, then returns separate transparent RGBA images that can be rearranged in a conventional editor. Those outputs give downstream tools individual visual elements to manipulate, which supports the release’s screenshot-to-code and image-to-PowerPoint workflows.

The Crello results currently come from inclusionAI’s evaluation. Independent reproduction will need the same dataset split, layer plans, hardware, precision, output resolution and timing method. The reported 4.3× speed advantage applies only to the named 20B baseline under that configuration.

Benchmark scope and deployment limits

• The leading rank is category-specific. Ming-Image leads open-weight models on the UI/UX board while ranking 45th on the general Text to Image board.
• The layer benchmark is vendor-reported. Its accuracy and speed claims await independent reproduction.
• Deployment requires high-memory datacenter hardware. The published configuration does not target laptops or typical workstation GPUs.
• Hosted access is unavailable. Teams currently need to operate the model and supporting infrastructure themselves.
• The reference settings are prescriptive. Changes to resolution, step count, guidance, precision or quantisation should be benchmarked for quality, latency and memory use.
• Leaderboard positions can move. Elo scores change as new models, prompts and preference votes enter the arena.

For developers building design agents, Ming-Image’s practical contribution is a pipeline that carries generated graphics into transparent assets, editable layers, interface code and presentation files. Its current deployment profile fits self-hosted experiments on datacenter GPUs, particularly for dashboards, slides, infographics and interface mockups where layout and legible text drive the result.