I use diffusion models, with fixed trained likelihood-based models, to achieve better sample quality than state-of-the-art gans. The improved architecture is sufficient to achieve this on unconditional image generation tasks, and classifier guidance techniques allow us to do so on class-conditional tasks. In the latter case, I find that scales to adjust classifier gradients to trade off diversity and fidelity. These guided diffusion models can reduce the sampling time gap between gans and diffusion models, al though with diffusion models, the model still requires mul tiple forward passes during sampling. Finally, by incorporating guidance via upsampling, I can further improve sample quality for high-resolution conditional image synthesis.
Created by LYCHEN(YUCHEN LI), 2024
Shenzhen, Guangdong, China
Shenzhen, Guangdong, China
Please contact me
ARTWORKS
EXHIBITIONS