AIBRIEF
Chinese design software vendor 兔展智能 (RabbitPre) launched a layer aware AI image tool on 2026 08 05, joining Alibaba's Qwen team and an arXiv preprint in a multi lab race to make generative models editable at the element level.
Ask an AI image generator to move a logo two inches to the left, swap the headline font, and export a transparent PNG for an e-commerce listing. The whole picture re-renders, and the designer is back to fixing it by hand. That workflow gap, not raw image quality, is the real bottleneck in generative design today, and a small group of model labs are now publicly racing to fix it.
On 2026-08-05, 兔展智能 (RabbitPre), a Chinese design-software company, launched RabbitVis, a production tool built on the company's in-house UniWorld-Design model. Unlike most image generators, which output a single flat image, UniWorld-Design is trained to produce and edit image layers separately, the way a Photoshop file does. The launch coverage frames layer-aware generation as the missing "second half" of AI image work: the part that has to ship inside real design pipelines rather than stay inside a prompt box.
The company also released a technical report on the model, and the self-reported I2L image-layering scores (RGB L1 of 0.1264, Alpha Soft IoU of 0.7325) are now part of the public comparison set. The numbers come from the vendor's own benchmarks and have not been independently reproduced.
The launch does not stand alone. Alibaba's Qwen team published Qwen-Image-Layered as an open-source project the same week, and a related preprint on layer decomposition is also on arXiv. The "editable layers" problem is now being chased in parallel across at least three labs.