TACO: Trajectory-Aware Curvature Observation for Memorization Localization and Mitigation in Diffusion Models

概要

Text-to-image diffusion models can reproduce memorized training content within localized image regions, making precise localization essential for targeted mitigation. However, curvature observations at a single denoising state can contain transient responses, while spatially uniform mitigation may disrupt prompt-relevant content. We present Trajectory-Aware Curvature Observation (TACO), a unified framework for local memorization detection and mitigation in diffusion models. Our empirical analysis reveals that memorization-related curvature becomes pronounced during late-stage denoising and persists at consistent spatial locations across neighboring states. Motivated by these observations, TACO aggregates prompt-dependent curvature responses along the late denoising trajectory and applies latent-guided spatial refinement to improve spatial coherence. The resulting memorization map serves both as a localization signal and as a spatial control signal for mitigation. Specifically, TACO uses this map to steer existing inference-time interventions toward memorization-related regions through spatial reweighting and localized guidance correction. Extensive experiments demonstrate that TACO consistently outperforms competing localization methods, and improves the Pareto trade-off between memorization suppression, semantic alignment, and perceptual quality. These results establish trajectory-aware spatial analysis as an effective link between memorization localization and targeted mitigation.

参考リンク

• Paper ↗

リンク先のライセンス・再現性・データ条件をご確認ください。自動分類や実装の目安は保証ではありません。

実装を始めるプロンプト

参考リンクを確認したうえで、次の内容をコピーして利用できます。

# 論文・記事
TACO: Trajectory-Aware Curvature Observation for Memorization Localization and Mitigation in Diffusion Models

# 概要
Text-to-image diffusion models can reproduce memorized training content within localized image regions, making precise localization essential for targeted mitigation. However, curvature observations…

# 原典
https://arxiv.org/abs/2610.41520

上記の手法を実装してください。実行前に依存関係とライセンスを確認してください。