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

UniWorld-Design reframes image generation around semantic RGBA layers rather than flattened pixels. It includes T2RGBA, which generates standalone transparent assets from text, and I2L, which converts a finished image into ordered, complete semantic layers using global instructions and per-layer prompts. The interface supports top-level decomposition, recursive decomposition, and targeted extraction. According to the abstract, I2L reduces per-layer RGB L1 error by 37% and delivers a 34% relative improvement in Alpha Soft IoU over Qwen-Image-Layered on Crello; T2RGBA also reports the highest CLIP Score among the named baselines.