Richard discusses the complexities of obtaining ground truth in image synthesis, highlighting the innovative use of customization techniques like dream booth. By training thousands of custom diffusion models, he and his team created a dataset that links training images to synthesized outputs, paving the way for more effective attribution analysis. However, he acknowledges that this is just the beginning of a challenging journey in disentangling the contributions of all training images.
