I am recently implementing DDPM model from scratch, and I discovered that UNet often tends to give noisy output in blank region. Here is an example with FashionMNIST, my DDPM seems to generate OK samples for clothes but not very good at generating blank background:

Generated FashionMNIST clothes with noisy background

Similarly, I had another project on lane detection with UNet, where I predict the sementic information for the lane-masks(lanes are left out in the mask, and we do binary classification on each pixel). UNet exhibit similar property, where the lane region is very noisy:

UNet lane detection

I found this property to be interesting, and my UNet implmentation for DDPM is not the same (DDPM uses ConvNext blocks and lane detection only uses Conv2Ds). Anyone know why?



You must log in to answer this question.