Hi @researcher171473,

The idea of using GANs or latent diffusion models to augment visual data instead of image mixing is indeed interesting. However, I have a few considerations:

1. Training GANs and diffusion models is typically more resource intensive than simple image mixing.
2. Ensuring that the generated examples are sufficiently informative and diverse to improve the classifier may require additional mechanisms (diversity regularization, adversarial training, latent space manipulation, domain-specific constraints, etc.).
3. The generated examples must retain their original semantics and class membership to effectively complement the training data.
4. The classifier may overfit the generated examples and lose performance on real data.

Despite these potential challenges, combining image blending with generative models could potentially yield better results. For example, GANs could be used to generate additional realistic samples that can then be mixed to increase diversity.