UniWorld-Design: From Pixel Generation to Layer-Native Design

2026-08-04

AI Summary

This work addresses the limitations of existing image generation methods, which are predominantly confined to pixel-level synthesis and struggle to produce structured, editable semantic content. To overcome this, the authors propose UniWorld-Design, a novel framework that unifies generation and editing within a native semantic RGBA layer space, treating layers as fundamental units for text- or image-to-structured-visual-content synthesis. The framework comprises two core models—Text-to-RGBA and Image-to-Layer—that integrate multimodal generation, semantic layer modeling, and instruction-conditioned control to enable instruction-driven hierarchical decomposition and object-level editing. Experiments on the Crello benchmark demonstrate that Image-to-Layer reduces RGB L1 error by 37% and improves Alpha Soft IoU by 34%, while Text-to-RGBA achieves significantly higher CLIP Scores than LayerDiffuse and OmniAlpha, confirming the approach’s superiority in structural semantic integrity and reusability.

Abstract

We introduce UniWorld-Design, a framework that redefines image generation from flat pixel synthesis to structured visual composition, with semantic RGBA layers as the atomic units of generation, understanding, and editing. Our key insight is that pixels define how an image is rendered, whereas layers define how an image is created, understood, and edited. Just as human designers create and manipulate visual content through layers rather than raw pixels, UniWorld-Design equips multimodal generative models with a layer-native design space. UniWorld-Design comprises two models. The Text-to-RGBA (T2RGBA) model generates standalone RGBA assets directly from text. The Image-to-Layer (I2L) model conditions on a finished image, a global instruction and per-layer prompts, and jointly produces ordered, complete semantic RGBA layers. Its instruction interface supports top-level decomposition, recursive decomposition and targeted extraction, making layering an instruction-addressable operation for agentic editing. Because I2L learns complete semantic objects rather than visible-pixel partitions, its layers stay usable when moved or removed. On the Crello benchmark, I2L reduces per-layer RGB L1 error by 37% and achieves a 34% relative improvement in Alpha Soft IoU over Qwen-Image-Layered. Separately, T2RGBA achieves the highest CLIP Score, outperforming LayerDiffuse and OmniAlpha.

Problem

• image generation
• semantic layers
• RGBA layers
• layer-native design
• structured visual composition

Innovation

• layer-native generation
• semantic RGBA layers
• instruction-addressable editing
• Text-to-RGBA
• Image-to-Layer