DeltaAug: A Directional Semantic Data Augmentation Framework Guided by Differential Reasoning

High-quality and semantically diverse training data remains a critical bottleneck in modern machine learning. Generative data augmentation (GDA) offers a potential solution; yet, existing approaches often face a fundamental tension: introducing meaningful semantic variation while preserving the core identity of the original sample. We introduce Directional Semantic Data Augmentation (DSDA), a conceptual framework that generates data along controllable semantic directions without disrupting the global context. To instantiate this paradigm, we propose DeltaAug, a framework that incorporates an explicit semantic difference modeling mechanism. By modeling the difference between the source and target states and injecting it as a directional control signal, DeltaAug enables fine-grained, interpretable augmentation within the diffusion process. Experiments across two visual benchmarks show that DeltaAug produces high-quality samples with stronger directional alignment (CLIP-T) and well-preserved visual consistency (DINO, CLIP-I). Integrating these augmentations further yields substantial improvements in both standard and fine-grained few-shot classification tasks, consistently outperforming conventional and recent generative methods.