1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization

TL;DR

1S-DAug addresses the challenge of robust few-shot generalization by providing a one-shot, test-time augmentation operator that generates diverse yet faithful variants from a single image using shape tweaks, controlled noise, and image-conditioned diffusion. It is model-agnostic and requires no parameter updates, delivering consistent improvements across four benchmarks and multiple backbones. The work couples empirical gains with theoretical analyses showing that diversity and radius reduction tighten margin-based generalization bounds, explaining the observed robustness to distribution shift. Practically, 1S-DAug offers a scalable data-side augmentation that enhances reliability in real-world, low-label settings without retraining, making it attractive for deployment in safety-critical perception systems.

Abstract

Few-shot learning (FSL) challenges model generalization to novel classes based on just a few shots of labeled examples, a testbed where traditional test-time augmentations fail to be effective. We introduce 1S-DAug, a one-shot generative augmentation operator that synthesizes diverse yet faithful variants from just one example image at test time. 1S-DAug couples traditional geometric perturbations with controlled noise injection and a denoising diffusion process conditioned on the original image. The generated images are then encoded and aggregated, alongside the original image, into a combined representation for more robust FSL predictions. Integrated as a training-free model-agnostic plugin, 1S-DAug consistently improves FSL across standard benchmarks of 4 different datasets without any model parameter update, including achieving over 10% proportional accuracy improvement on the miniImagenet 5-way-1-shot benchmark. Codes will be released.

Figures (6)

• Figure 1: Pipeline of our one-shot, test-time augmentation. Given a query and supports, we apply a shape tweak and controlled noise, then perform attention-conditioned diffusion to synthesize class-faithful variants. Features from the original and the generated views are averaged before the few-shot head. For contrast, traditional test-time geometric transformations provide limited diversity.
• Figure 2: Effect of noise and conditioning. Qualitative ablation on a single input across increasing noise levels. Shape-only edits yield limited diversity; adding noise increases diversity but may reduce fidelity without conditioning. Attention-conditioned diffusion preserves class-defining content while enabling controlled pose/appearance changes; excessive noise without the image condition degrades faithfulness.
• Figure 3: Effect of noise and shape tweak. Comparison across three settings: no shape tweak, shape tweak only, and shape + noise + attention-conditioned diffusion (ours). Increasing noise and including shape tweak expand diversity, and our full setting provides the best balance for both diversity and faithfulness.
• Figure 4: Failure modes of GAN-based image-to-image translation. Examples where image-to-image GAN translation fails to preserve the intended class. Rows contain target class and failed GAN outputs with typical artifacts.
• Figure 5: More qualitative results from 1S-DAug. Each pair contains the original image followed by our synthesis. All visualization pairs are random without cherry-picking.

Theorems & Definitions (13)

• Proposition 5.1: Pairwise Risk Decomposition
• Theorem 5.2: Encoder margin bound
• Lemma 5.3: Rademacher control via radius
• Proposition 5.4: Probabilistic radius reduction with $M$ augmentations
• Theorem 2.1: Restatement of Theorem \ref{['thm:encoder-margin']}
• Lemma 2.2: Rademacher complexity of the encoder score class
• proof : Proof sketch
• Lemma 2.3: Empirical margin stability
• proof
• Proposition 2.4: Radius reduction with $M$ augmentations