Diffusemix: Label-Preserving Data Augmentation with Diffusion Models

Target (T) CutMix (S+T) Mixup (S+T) GridMix (S+T) ResizeMix (S+T) DiffuseMix (S) AugMix (T) PixMix (T) IPMix (T) PuzzleMix (S+T) DiffuseMix (S) SmoothMix (S+T) DiffuseMix (T) AdaAutoMix (S+T) DiffuseMix (T) Figure 1. Top row: existing mixup methods interpolate two different training images [22, 49]. Bottom row: label-preserving methods. For each input image, DIFFUSEMIX employs conditional prompts to obtain generated images. The input image is then concatenated with a generated image to obtain a hybrid image. Each hybrid image is blended with a random fractal to obtain the final training image.