MaskAttn-SDXL: Controllable Region-Level Text-To-Image Generation

Abstract:Diffusion models have achieved strong results in text-to-image generation, but important limitations remain as prompts become more structured and multi-object. On the architecture side, U-Net backbones are efficient and stable, yet their locality makes global coordination harder, while Transformer-based diffusion models improve global interactions but at substantially higher compute and memory cost. In parallel, compositional reliability remains weak: models often mix attributes across objects, violate spatial relations, or omit requested entities, and these errors are not reliably reflected by global metrics such as FID or CLIP-based scores. To address these issues without changing the SDXL pipeline, we propose MaskAttn-SDXL, a plug-in module that injects token-conditioned spatial gating into cross-attention logits before softmax. The gating sparsifies token-to-location interactions to suppress irrelevant bindings while preserving the pretrained backbone and standard sampling process, requiring no external supervision or inference-time editing.

Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)

arXiv:2509.15357 [cs.CV]

(or arXiv:2509.15357v3 [cs.CV] for this version)