I have a question about the sampler you employed in your work. You mentioned that you used the original DDPM sampler but set the variance to 0 to achieve better performance. However, such a setting is not theoretically equal to the original sampler. DDIM [1] is a commonly used diffusion sampler w/o variance and it can be proven that it has the same optimization target as DDPM. So, why do you choose to employ DDPM w/o variance instead of DDIM? Does this choice have any theoretical reasons? Or just an empirical choice?
Thank you again for your excellent work. Looking forward to your reply.
Zj-BinXia
commented
10 months ago
Thank you for your contribution to the community.
I have a question about the sampler you employed in your work. You mentioned that you used the original DDPM sampler but set the variance to 0 to achieve better performance. However, such a setting is not theoretically equal to the original sampler. DDIM [1] is a commonly used diffusion sampler w/o variance and it can be proven that it has the same optimization target as DDPM. So, why do you choose to employ DDPM w/o variance instead of DDIM? Does this choice have any theoretical reasons? Or just an empirical choice?
Thank you again for your excellent work. Looking forward to your reply.
[1] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising Diffusion Implicit Models", ICLR 2021.