Core Concepts
Perceptual loss improves diffusion model sample quality by incorporating implicit perceptual guidance.
Stats
"Diffusion models trained with mean squared error loss tend to generate unrealistic samples."
"Classifier-free guidance improves sample quality by removing the need for an external classifier network."
"Perceptual loss can be directly applied to diffusion training to improve sample quality."
Quotes
"Our proposed self-perceptual loss fixes this issue and can generate realistic samples without CFG."
"Perceptual loss can be directly applied to diffusion training to improve sample quality."
"Classifier-free guidance improves classifier guidance by removing the need for an external classifier network."