Localizing Knowledge in Diffusion Transformers
This work addresses the need for interpretability and controllability in generative models, particularly for researchers and practitioners using DiTs, though it is incremental as it extends prior localization methods from UNet-based to DiT-based architectures.
The paper tackled the problem of understanding how knowledge is distributed in Diffusion Transformer (DiT)-based models by proposing a method to localize specific knowledge types within DiT blocks, showing that identified blocks are interpretable and causally linked to generated outputs, and applying this to model personalization and knowledge unlearning with improved efficiency and performance.
Understanding how knowledge is distributed across the layers of generative models is crucial for improving interpretability, controllability, and adaptation. While prior work has explored knowledge localization in UNet-based architectures, Diffusion Transformer (DiT)-based models remain underexplored in this context. In this paper, we propose a model- and knowledge-agnostic method to localize where specific types of knowledge are encoded within the DiT blocks. We evaluate our method on state-of-the-art DiT-based models, including PixArt-alpha, FLUX, and SANA, across six diverse knowledge categories. We show that the identified blocks are both interpretable and causally linked to the expression of knowledge in generated outputs. Building on these insights, we apply our localization framework to two key applications: model personalization and knowledge unlearning. In both settings, our localized fine-tuning approach enables efficient and targeted updates, reducing computational cost, improving task-specific performance, and better preserving general model behavior with minimal interference to unrelated or surrounding content. Overall, our findings offer new insights into the internal structure of DiTs and introduce a practical pathway for more interpretable, efficient, and controllable model editing.