arXiv:2605.26756v2 Announce Type: replace 
Abstract: Diffusion models can unintentionally memorize training samples, raising concerns about privacy and copyright. While recent methods can detect memorization, they often rely on global or model-specific signals and provide limited insight into where memorization appears within a generated image. We provide a geometric characterization of local memorization as a coordinate-wise variance collapse. However, such collapse can also arise from intrinsic data constraints rather than overfitting. To isolate overfitting-driven memorization, we propose curvature-difference methods that subtract the curvature of an underfitted baseline, either the unconditional model or a less-trained version of itself. We further derive a score-difference proxy that provides a geometric explanation for the widely used score-difference-based detection metric. Experiments on Stable Diffusion, evaluated against ground-truth memorization masks, show that our method outperforms the prior attention-based localization method. Code is available at https://github.com/Gwangho99/mem-curv-diff.

أظهرت الأبحاث الحديثة أن نماذج الانتشار يمكن أن تتذكر عن غير قصد عينات التدريب، مما يثير مخاوف كبيرة بشأن الخصوصية وحقوق الطبع والنشر. تقدم الدراسة توصيفًا هندسيًا للذاكرة المحلية من خلال انهيار التباين حسب الإحداثيات، وتقترح طرق فرق الانحناء لعزل الذاكرة الناتجة عن الإفراط في التكيف.

Investigaciones recientes han identificado que los modelos de difusión pueden memorizar involuntariamente muestras de entrenamiento, lo que plantea preocupaciones significativas sobre la privacidad y los derechos de autor. El estudio introduce una caracterización geométrica de la memorización local a través del colapso de varianza por coordenadas, proponiendo métodos de diferencia de curvatura para aislar la memorización impulsada por el sobreajuste.

Des recherches récentes ont révélé que les modèles de diffusion peuvent mémoriser involontairement des échantillons d'entraînement, soulevant des préoccupations importantes concernant la vie privée et le droit d'auteur. L'étude introduit une caractérisation géométrique de la mémorisation locale par le biais d'un effondrement de variance coordonnée, proposant des méthodes de différence de courbure pour isoler la mémorisation due au surajustement.

Recent research has identified that diffusion models can inadvertently memorize training samples, raising significant concerns regarding privacy and copyright. The study introduces a geometric characterization of local memorization through coordinate-wise variance collapse, proposing curvature-difference methods to isolate overfitting-driven memorization.