Training-Free Concept Disentanglement in Text-to-Image Generation via Detection-Guided Attention Control

Large-scale Text-to-Image (T2I) diffusion models have achieved remarkable photorealism but frequently suffer from attribute leakage (also known as concept entanglement), where visual attributes such as color or texture bleed into unintended regions. Existing solutions often rely on costly fine-tuning or auxiliary adapters, limiting their flexibility. To address this, we propose a novel Training-Free Spatial-Gated Attention Control framework that enforces strict attribute disentanglement during the inference stage. Specifically, we leverage an open-vocabulary detector (OwlViT) to extract zero-shot layout priors and introduce a Spatial-Gated Boosting mechanism to dynamically intervene in Cross-Attention layers. This mechanism amplifies target signals within designated regions while suppressing background noise, ensuring precise concept localization. Quantitative evaluation demonstrates that our method achieves a 11% relative improvement in CLIP Score compared to traditional inpainting baselines, indicating superior semantic consistency. Furthermore, our approach exhibits an emergent "layout correction" capability, where it automatically refines unnatural bounding box constraints to align with physical plausibility (e.g., repositioning accessories), offering a robust and efficient solution for controllable generation. Furthermore, our approach exhibits an emergent "layout correction" capability. Our code and data are publicly available at https://github.com/QCYTSN/Training-Free-Concept-Disentanglement-in-T2I-Generation-via-Detection-Guided-Attention-Control.