DeLeaker: Dynamic Inference-Time Reweighting For Semantic Leakage Mitigation in Text-to-Image Models
This addresses a critical issue for users of text-to-image models by improving semantic precision, though it is an incremental advance as it builds on existing attention mechanisms.
The paper tackles the problem of semantic leakage in text-to-image models, where unintended feature transfer occurs between distinct entities, and introduces DeLeaker, an inference-time method that dynamically reweights attention maps to mitigate this issue, achieving consistent outperformance over baselines without compromising fidelity or quality.
Text-to-Image (T2I) models have advanced rapidly, yet they remain vulnerable to semantic leakage, the unintended transfer of semantically related features between distinct entities. Existing mitigation strategies are often optimization-based or dependent on external inputs. We introduce DeLeaker, a lightweight, optimization-free inference-time approach that mitigates leakage by directly intervening on the model's attention maps. Throughout the diffusion process, DeLeaker dynamically reweights attention maps to suppress excessive cross-entity interactions while strengthening the identity of each entity. To support systematic evaluation, we introduce SLIM (Semantic Leakage in IMages), the first dataset dedicated to semantic leakage, comprising 1,130 human-verified samples spanning diverse scenarios, together with a novel automatic evaluation framework. Experiments demonstrate that DeLeaker consistently outperforms all baselines, even when they are provided with external information, achieving effective leakage mitigation without compromising fidelity or quality. These results underscore the value of attention control and pave the way for more semantically precise T2I models.