search

IceClear / StableSR

[IJCV2024] Exploiting Diffusion Prior for Real-World Image Super-Resolution
https://iceclear.github.io/projects/stablesr/
Other
2.13k stars 138 forks source link

DDIM inference eta #108

Open salvadog opened 9 months ago

salvadog commented 9 months ago

Hi, thanks for your great job. I'm curious about the use of ddim sample with eta=1, which is equivalent to respaced DDPM. And when I tried to sample with ddim eta=0.0, which means no addtional stochastic noise, but the sample results were not good and full of random artifacts. Do you have any clues?

IceClear commented 9 months ago

Hi. Thanks for your interest of our work. I used to test on eta=0.0 and did not see too much quality degradation. You may try to use different seeds to check if the artifacts always exist. I did observe that the performance of eta=0.0 is not as good as eta=1.0 because of the lack of randomness. But the performance should not be too bad. You can show me the results if still does not work. Thanks.