jychoi118 / ilvr_adm

ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models (ICCV 2021 Oral)
MIT License
417 stars 51 forks source link

Why not use L2 distance gradient? #31

Open GiilDe opened 1 year ago

GiilDe commented 1 year ago

Qurious as to why you use $\phi(y{t-1}) - \phi(x{t-1})$ and not use $\nabla{x{t-1}}||\phi(y{t-1}) - \phi(x{t-1})||_2^2$ similarly to classifier guidance?