improved-diffusion
improved-diffusion copied to clipboard
Experiments on extrapolation with diffusion
Goal
We're trying to answer the question, can diffusion models denoise beyond the original level of noise in the dataset?
Results
Extrapolating the time step past 0 doesn't seem to work on CIFAR-10.
Need to check whether this is something hardcoded due to e.g. alpha, beta scheduling.
Time step 0:

Time step -1:

It may be easier to start with a synthetic dataset and go from there.