TMLR Journal 2025 Journal Article
Attention Overlap Is Responsible for The Entity Missing Problem in Text-to-image Diffusion Models!
- Arash Mari Oriyad
- Mohammadali Banayeeanzade
- Reza Abbasi
- Mohammad Hossein Rohban
- Mahdieh Soleymani Baghshah
Text-to-image diffusion models such as Stable Diffusion and DALL-E have exhibited impressive capabilities in producing high-quality, diverse, and realistic images based on textual prompts. Nevertheless, a common issue arises where these models encounter difficulties in faithfully generating every entity specified in the prompt, leading to a recognized challenge known as entity missing in visual compositional generation. While previous studies indicated that actively adjusting cross-attention maps during inference could potentially resolve the issue, there has been a lack of systematic investigation into the specific objective function required for this task. In this work, we thoroughly investigate three potential causes of entity missing from the perspective of cross-attention maps: insufficient attention intensity, excessive attention spread, and significant overlap between attention maps of different entities. Through comprehensive empirical analysis, we found that optimizing metrics that quantify the overlap between attention maps of entities is highly effective at mitigating entity missing. We hypothesize that during the denoising process, entity-related tokens engage in a form of competition for attention toward specific regions through the cross-attention mechanism. This competition may result in the attention of a spatial location being divided among multiple tokens, leading to difficulties in accurately generating the entities associated with those tokens. Building on this insight, we propose four overlap-based loss functions that can be used to implicitly manipulate the latent embeddings of the diffusion model during inference: Intersection over union (IoU), center-of-mass (CoM) distance, Kullback–Leibler (KL) divergence, and clustering compactness (CC). Extensive experiments on a diverse set of prompts demonstrate that our proposed training-free methods substantially outperform previous approaches on a range of compositional alignment metrics, including visual question-answering, captioning score, CLIP similarity, and human evaluation. Notably, our method outperforms the best baseline by $9\%$ in human evaluation.