Breaking the Synthetic-Real Domain Shortcut for Training-Free Generative Replay-based Class Incremental Learning

2026-07-24

AI Summary

This work addresses a critical yet overlooked issue in zero-shot class-incremental learning: the tendency of models to exploit domain-specific cues—arising from directly mixing synthetic old-class samples with real new-class data—rather than genuine semantic features, leading to performance degradation. To tackle this synthetic-to-real domain shortcut problem in generative replay, the authors propose DREAM, a novel approach that leverages a training-free text-to-image generator to synthesize old-class exemplars and introduces subspace correction with orthogonal projection to eliminate domain bias. Additionally, real-anchored prototype regularization is incorporated to enhance semantic alignment. Requiring neither storage of real samples nor additional training, DREAM achieves state-of-the-art performance across four benchmark datasets, significantly outperforming existing zero-shot class-incremental learning methods.

Abstract

Class-incremental learning (CIL) requires models to continuously acquire new knowledge while avoiding catastrophic forgetting. While exemplar replay is effective, it raises concerns regarding privacy and storage. Thus, generative replay has emerged as a viable alternative, synthesizing old data using frozen pretrained text-to-image (T2I) models without any extra training. However, we observe that directly mixing synthetic old-class data with real new-class data during incremental training leads to significant performance degradation. This issue stems from a "domain shortcut", where models rely on domain-discriminative features instead of semantic class cues. To address this, we propose DREAM ($\underline{\mathbf{D}}$omain-$\underline{\mathbf{R}}$egularized $\underline{\mathbf{E}}$xemplar-free $\underline{\mathbf{A}}$lignment $\underline{\mathbf{M}}$odel), which uses a training-free generator to synthesize old-class data and eliminates domain shortcut via subspace rectification and orthogonal projection, while reinforcing semantic alignment through real-anchored prototype regularization. Extensive experiments on 4 datasets demonstrate that DREAM outperforms existing exemplar-free CIL methods and achieves state-of-the-art performance. Our source code is available at https://github.com/Light-ZhangTao/DREAM.

Problem

• class-incremental learning
• generative replay
• domain shortcut
• catastrophic forgetting
• synthetic-real domain gap

Innovation

• generative replay
• class-incremental learning
• domain shortcut
• training-free
• semantic alignment