3 ms·

I also noticed that they present it as if other diffusion models were not invertible, but they are. Also, they appear to be using DDPM++, the exact same neural

I also noticed that they present it as if other diffusion models were not invertible, but they are. Also, they appear to be using DDPM++, the exact same neural network architecture as Stable Diffusion? Article says: "to train the neural network, which in this case is a U-Net (DDPM++ backbone)." Also it's kinda weird that they show videos generating images from noise (like Imagen / Stable Diffusion) but don't cite any recent diffusion paper. EDIT: In fact they don't cite ANY 2022 paper. So my guess would be that they submitted this last year and now it's been made public because now the NeurIPS 2022 conference is taking place. But most likely, the actual research here predates the Stable Diffusion release.

SleekEagle 4y ago
There is a stochastic element to diffusion models which is not present in PFGMs. Further, recall that Diffusion Models and PFGMs alike are "metamodels" - mathematical frameworks in which any function approximator can, in theory, be placed. They are not an architecture. In this case, the authors use the DDPM++ backbone which is the specific function approximator that they implement in their PFGM framework.