← Back to Computer Vision
cs.CV

How to generate editable layers like Photoshop does

Zhicong Tang, Zhao Zhang, Jingye Chen, Mohan Zhou, Yifan Pu, Yuchi Liu, Yalong Bai, Ethan Smith, Yuhui Yuan

May 26, 2026

Most image generators produce flat pictures; this one outputs editable layers (text, shapes, backgrounds) you can rearrange like in Photoshop. MRT, a 20-billion-parameter model trained on 10M design samples, handles three workflows—text-to-layers, image-to-layers, and layer-to-layer editing—in one framework. An overflow-aware canvas lets layers extend beyond boundaries. Distilled to 8 steps, it runs in real time on consumer hardware while using 50–90% less GPU memory than concurrent systems.
Published as MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale arXiv:2605.27235
Read the original paper →