Mitigating Memorization in Text-to-Image Diffusion via Region-Aware Prompt Augmentation and Multimodal Copy Detection

TL;DR

This paper introduces two novel methods, RAPTA and ADMCD, to reduce memorization and improve copyright detection in text-to-image diffusion models, balancing diversity, fidelity, and detection accuracy.

Contribution

The paper presents Region-Aware Prompt Augmentation and Attention-Driven Multimodal Copy Detection as new techniques to mitigate memorization and detect copying without extensive annotated datasets.

Findings

RAPTA reduces overfitting while maintaining image quality.

ADMCD effectively detects copying with high reliability.

Methods outperform existing single-modal metrics.

Abstract

State-of-the-art text-to-image diffusion models can produce impressive visuals but may memorize and reproduce training images, creating copyright and privacy risks. Existing prompt perturbations applied at inference time, such as random token insertion or embedding noise, may lower copying but often harm image-prompt alignment and overall fidelity. To address this, we introduce two complementary methods. First, Region-Aware Prompt Augmentation (RAPTA) uses an object detector to find salient regions and turn them into semantically grounded prompt variants, which are randomly sampled during training to increase diversity, while maintaining semantic alignment. Second, Attention-Driven Multimodal Copy Detection (ADMCD) aggregates local patch, global semantic, and texture cues with a lightweight transformer to produce a fused representation, and applies simple thresholded decision rules to detect copying without training with large annotated datasets. Experiments show that RAPTA reduces overfitting while maintaining high synthesis quality, and that ADMCD reliably detects copying, outperforming single-modal metrics.