MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale

Researchers introduced MRT, a 20B-parameter masked region diffusion model for layered image generation and editing.

It unifies text-to-layers, image-to-layers, and layers-to-layers tasks using over 10 million multilingual design samples.

MRT adds an overflow-aware canvas for better boundary handling and uses diffusion distillation to enable 8-step real-time generation.

The model uses less memory and delivers better quality and speed than prior research and commercial systems.