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

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chang, Yu, Chen, Jiahao, Cheng, Anzhe, Bogdan, Paul
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916017006968832
author Chang, Yu
Chen, Jiahao
Cheng, Anzhe
Bogdan, Paul
author_facet Chang, Yu
Chen, Jiahao
Cheng, Anzhe
Bogdan, Paul
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2509_15357
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MaskAttn-SDXL: Controllable Region-Level Text-To-Image Generation
Chang, Yu
Chen, Jiahao
Cheng, Anzhe
Bogdan, Paul
Computer Vision and Pattern Recognition
Machine Learning
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.
title MaskAttn-SDXL: Controllable Region-Level Text-To-Image Generation
topic Computer Vision and Pattern Recognition
Machine Learning
url https://arxiv.org/abs/2509.15357