Diffusion models are trained by reproducing training data and the "only" uniqueness comes from interpolating between whatever space exists inbetween. Memorization is negative for an LLM but not for a diffusion model.
This has also reached the courts where Diffusion AIs have even been found to be infringing. LLMs are solving unique challenges (in maths, as an example) while diffusion models still have a hard time showing that they're not just copying.
https://alphasignal.ai/news/google-research-finally-explains-why-diffusion-models-don-t-just-copy-training
mbpaz (npub1xcd…hnp8) Chad McCullough (npub1ded…kves) elementary (npub16nm…cc46)
