Questions for the seminar paper "Extracting Training Data from Diffusion Models (https://arxiv.org/pdf/2301.13188)".
-----------------------------------------------------------------------------------------------
Please send your answers to: sunagadb@cs.uni-freiburg.de

1. What factors influence whether an image will be memorized by a diffusion model? (2-3 sentences)
2. Why do the authors define a new metric ("tile based L2") for image memorization instead of using traditional L2 distance alone?(1-2 sentences)
3. Why is timestep selection critical in membership inference attacks on diffusion models, and how do the authors optimize it?(2-3 sentences)