Questions for the seminar paper "Boosting latent diffusion wiht flow matching"
-----------------------------------------------------------------------------------------------
Please send your answers to: dienertj@cs.uni-freiburg.de

1) What's the author's argument for splitting the diffusion process into a part with low- & high stochasticity?
2) What is 'noise augmentation', and where is it applied in the pipeline?
3) The authors observe degraded performance when sampling from a stable diffusion v1.5 model at lower resolutions than it was trained on. What do they do to mitigate this?