Random horizontal flipping during training improves sample quality in image diffusion models

The DDPM paper found that applying random horizontal flips to training images slightly improved sample quality for CIFAR-10. This is a simple data augmentation that effectively doubles the training data by presenting mirror-image versions of each sample. For the Fashion-MNIST tutorial, horizontal flips are also applied via transforms.RandomHorizontalFlip(). This is an example of domain-appropriate augmentation: horizontal flips are realistic (a T-shirt looks the same mirrored), unlike vertical flips which would be unrealistic for clothing or natural scenes.

Examples

PyTorch transform: transforms.RandomHorizontalFlip(). Applied within dataset.with_transform() so augmentation happens on-the-fly during training, not stored.

Assessment

Explain why horizontal flip augmentation is appropriate for CIFAR-10 and Fashion-MNIST but might not be appropriate for other datasets. How does on-the-fly augmentation differ from pre-augmenting the dataset?