Training ESRGAN, I quickly learned that AI sniffs artifacts humans can't see. And I got particularly funky results training a model on selected previous generat

Training ESRGAN, I quickly learned that AI sniffs artifacts humans can't see. And I got particularly funky results training a model on selected previous generation upscales. Hence I disagree. Even of the training data curators are perfect, there is a risk of "model inbreeding." And the curators are not perfect. In fact, they seem to prioritize (by necessity) volume and speed over quality and curation when training these LLMs and diffusion models.