DeLeaker: Dynamic Inference-Time Reweighting For Semantic Leakage Mitigation in Text-to-Image Models
Mor Ventura, Michael Toker, Or Patashnik, Yonatan Belinkov, Roi Reichart
2025-10-23
Summary
This paper focuses on a problem with AI image generators, specifically that they sometimes accidentally blend features from different things you ask for into a single image. It introduces a new method called DeLeaker to fix this issue and also provides a new dataset to help researchers test these kinds of problems.
What's the problem?
AI models that create images from text are getting really good, but they can still 'leak' information between different objects in the image. Imagine asking for a picture of a red car and a blue truck, but the AI accidentally gives the car some blue features or the truck some red ones. This happens because the AI isn't perfectly separating the concepts of 'car' and 'truck' during image creation, and existing solutions often require a lot of extra work or information.
What's the solution?
The researchers developed DeLeaker, which works by subtly changing how the AI pays 'attention' to different parts of the image while it's being generated. It doesn't require any extra training or information; it works during the image creation process itself. DeLeaker focuses the AI on keeping the features of each object separate and distinct, reducing the unwanted blending. They also created a new dataset, SLIM, with over a thousand images specifically designed to test for this leakage problem, and a way to automatically measure how well different methods perform.
Why it matters?
This work is important because it improves the accuracy and control of AI image generation. By preventing unwanted feature mixing, DeLeaker helps create images that more closely match what the user intended. This is a step towards more reliable and precise AI tools for creative tasks, and the new dataset will help other researchers develop even better solutions in the future.
Abstract
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.