arXiv:2511.16870v3 Announce Type: replace-cross 
Abstract: Enforcing alignment between the internal representations of diffusion or flow-based generative models and those of pretrained self-supervised encoders has recently been shown to provide a powerful inductive bias, improving both convergence and sample quality. In this work, we extend this idea to inverse problems, where pretrained generative models are employed as priors. We propose applying representation alignment (REPA) between diffusion or flow-based models and a DINOv2 visual encoder, to guide the reconstruction process at inference time. Although ground-truth signals are unavailable in inverse problems, we empirically show that aligning model representations of approximate target features can substantially enhance reconstruction quality and perceptual realism. We provide theoretical results showing (a) that REPA regularization can be viewed as a variational approach for minimizing a divergence measure in the DINOv2 embedding space, and (b) how under certain regularity assumptions REPA updates steer the latent diffusion states toward those of the clean image. These results offer insights into the role of REPA in improving perceptual fidelity. Finally, we demonstrate the generality of our approach by We integrate REPA into multiple state-of-the-art inverse problem solvers, and provide extensive experiments on super-resolution, box inpainting, Gaussian deblurring, and motion deblurring confirming that our method consistently improves reconstruction quality, while also providing efficiency gains reducing the number of required discretization steps.

أدت التطورات الأخيرة في نمذجة الجيل إلى اقتراح محاذاة التمثيل (REPA) بين نماذج الانتشار أو النماذج المعتمدة على التدفق ومشفّرات الصور DINOv2، بهدف تحسين عملية إعادة البناء في المشكلات العكسية حيث لا توجد إشارات حقيقية. تُظهر هذه الطريقة أن محاذاة تمثيلات النماذج يمكن أن تحسن بشكل كبير من جودة إعادة البناء والواقعية الإدراكية.

Los recientes avances en modelado generativo han llevado a la propuesta de alineación de representaciones (REPA) entre modelos de difusión o basados en flujo y codificadores visuales DINOv2, con el objetivo de mejorar el proceso de reconstrucción en problemas inversos donde no hay señales de verdad de terreno. Este enfoque demuestra que alinear las representaciones de los modelos puede mejorar significativamente la calidad de la reconstrucción y el realismo perceptual.

Des avancées récentes dans la modélisation générative ont conduit à la proposition d'un alignement de représentation (REPA) entre les modèles de diffusion ou basés sur le flux et les encodeurs visuels DINOv2, visant à améliorer le processus de reconstruction dans les problèmes inverses où les signaux de vérité de terrain sont absents. Cette approche démontre que l'alignement des représentations des modèles peut considérablement améliorer la qualité de reconstruction et le réalisme perceptuel.

Recent advancements in generative modeling have led to the proposal of representation alignment (REPA) between diffusion or flow-based models and DINOv2 visual encoders, aimed at enhancing the reconstruction process in inverse problems where ground-truth signals are absent. This approach demonstrates that aligning model representations can significantly improve reconstruction quality and perceptual realism.