improved-diffusion
improved-diffusion copied to clipboard
Why white or black image get the lowest bpd?
When I use a pure white image as data to run scripts/image_nll.py, I get the lowest bpd 0.059...(and black 0.096...). But a normal image in CIFAR10 only gets 3.2 or so. (4000 steps, model: cifar10_uncond_50M_500K.pt) Why?
When you think intuitively about it, that makes sense. Bits-per-dimension metric tells us how many bits per dimension (H x W x C of image) do we need to encode the image. I would expect low bpd for white and black images, due to their simplicity.