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

2026-05-26

AI Summary

This work addresses the limited capability of existing methods in large-scale hierarchical image generation and editing, particularly their inability to support layer reuse, manipulation, and composition. To overcome these challenges, we propose MRT—a 20-billion-parameter masked region diffusion model that unifies text-to-layer, image-to-layer, and layer-to-layer tasks within a single shared framework. MRT introduces a novel masked region Transformer architecture, an overflow-aware canvas layer to handle boundary inconsistencies and semi-transparent backgrounds, and leverages diffusion distillation for accelerated inference. Experimental results demonstrate that MRT significantly outperforms current state-of-the-art approaches and commercial systems across all three tasks: it achieves higher image-to-layer quality than Qwen-Image-Layered while offering 10–100× faster inference and reducing GPU memory consumption by 50–90%.

Abstract

Layered image generation and editing is a fundamental capability that enables layer-wise reuse, editing, and composition of generated visual content, analogous to word-level editing in natural language. Despite its importance, this remains an underexplored area at scale. To address this gap, we present MRT, a 20B-parameter masked region diffusion model tailored for multi-layer transparent image generation and editing, trained on over 10M multilingual design samples spanning diverse aspect ratios and textual prompts. To fully leverage this scale, we make two key technical contributions. First, we unify three complementary tasks including text-to-layers, image-to-layers, and layers-to-layers within a shared masked region diffusion framework, where selective token masking enables flexible layer-wise generation and editing. Second, to enable overflow layer generation, we introduce an overflow-aware canvas layer that handles boundary inconsistencies and supports semi-transparent background synthesis, enabling complete editable layers extending beyond visible canvas boundaries. Additionally, we apply diffusion distillation to achieve 8-step, real-time multi-layer generation with minimal quality degradation. Extensive experiments demonstrate that our framework substantially outperforms prior state-of-the-art approaches, including various commercial systems, across all three tasks, establishing a new benchmark for multi-layer transparent image generation. Notably, our model significantly outperforms the concurrent Qwen-Image-Layered model in image-to-layers quality according to user-study results, while achieving 10-100\times faster inference and reducing activation GPU memory consumption by 50-90\% during image-to-layer inference.

Problem

• layered image generation
• image editing
• multi-layer transparency
• scalable generation
• visual content composition

Innovation

• Masked Region Diffusion
• Layered Image Generation
• Overflow-aware Canvas
• Diffusion Distillation
• Multi-layer Editing