Martin Uecker on Nostr: Your link seems to say just the opposite. (BTW: I do research on diffusion models.) ...

Your link seems to say just the opposite. (BTW: I do research on diffusion models.) The idea that diffusion models reproduce training data is based on the (incorrect) idea that they would exactly learn the score, which is generally not true. Although by the same argument one could argue that an LLM trained to predict the next token also learns reproducing training data.