Image retrieval outperforms diffusion models on data augmentation

• DOI: 10.48550/arxiv.2304.10253
• Publication Date: 2023-01-01
• Authors: Max F. Burg; Florian Wenzel; Dominik Zietlow; Max Horn; Osama Makansi; Francesco Locatello; Chris Russell
• Keywords: Training set; Baseline (sea)

Many approaches have been proposed to use diffusion models augment training datasets for downstream tasks, such as classification. However, are themselves trained on large datasets, often with noisy annotations, and it remains an open question which extent these contribute classification performance. In particular, unclear if they generalize enough improve over directly using the additional data of their pre-training process augmentation. We systematically evaluate a range existing methods generate images from study new extensions assess benefit Personalizing towards target outperforms simpler prompting strategies. model alone, via simple nearest-neighbor retrieval procedure, leads even stronger Our explores potential in generating data, surprisingly finds that sophisticated not yet able beat strong image baseline vision tasks.