Denoising Diffusion models from first principle in Julia

Re:Denoising Diffusion models from first principle in Julia by radarsat1 on Saturday January 07, 2023 @12:20PM (#34290235)

Thanks for the corrections, I was including autoencoders that use an additional adversarial loss (such as VQGAN) when I said GAN.

> Dalle 1 works thx to the autoregressive model (also no GAN)

It uses an autoregressive model to predict codes for a pretrained VQGAN, doesn't it?

Doesn't Stable Diffusion's autoencoder also use an adversarial loss? Otherwise wouldn't it suffer the typical blurring problems well known to MSE?

[ Parent | on HN ]

Re:Denoising Diffusion models from first principle in Julia by GaggiX on Saturday January 07, 2023 @01:39PM (#34291089)

Yes, all the autoencoders you see used in practice have adversarial loss + MSE + perceptual loss, the VAE used with Stable Diffusion also uses KL regularization, while VQGAN uses all other losses to make use of the codebook.

[ Parent | on HN ]