FinnBehrendt/patched-Diffusion-Models-UAD

loss function

khanhpham3799 opened this issue · 1 comments

Hi, may I ask what loss functions you have been using while training diffusion model? Did you include variational lower bound? It could be great if you can implement the function in detail. Many thanks!

Hi, sorry for the late response. The implemented losses are l1 and l2 loss computed across the input image and the reconstruction. Throughout our experiments, we utilized the l1 loss. We do not directly use the VLB as loss but follow the approximation that is also shown in the original paper of Ho et al. with the distinction that we do not predict the noise but the actual reconstruction.
I hope this helps!
Best Finn