You're right. I was confusing them. The disconnect between the discussion and approach to training music models and image models still applies, although the hypocrisy I implied does not. Harmonai are approaching it sensibly. Huggingface and most other image diffusion models developers are not.