search

adobe-research / SuperGaussian

ECCV 2024 SuperGaussian for generic 3D upsampling
https://supergaussian.github.io/
Other
27 stars 0 forks source link

Is the pre-trained weight of RealBasicVSR fine-tuned one? #4

Open DHPark98 opened 1 week ago

DHPark98 commented 1 week ago

Thank you for your great work. I have a simple question about RealBasicVSR which is provided in the README.md.

Is the model weight fine-tuned one, or the pre-trained one from its original repo?

yshen47 commented 1 week ago

Thanks for follow-up! No, the weight of RealBasicVSR is not finetuned. It is the same one from the originial repo. Meanwhile, if you wanna compare to your own dataset, I'd recommend trying to plug in Upscale-a-video as video prior in SuperGaussian, which could give better results than RealBasicVSR. They recently released checkpoint and codebase.

QianyiWu commented 1 day ago

Hi @yshen47

Thanks for your work and open-source. I tried upscale-a-video prior in SuperGaussian. But the quantitative results are not as good as RealBasicVSR without fine-tuning. Do you have any suggestions about the hyper-parameters when using upscale-a-video?

The quantitative results I obtained are given as follow: LPIPS: 0.2518 NIQE: 9.2263 FID: 48.46

The hyper-parameters of Upscale-a-video are set as: -n 30, -g 1 -s 20.

BTW, I noticed higher noise level or higher CFG scale will leads to over-saturated artifacts and inconsistency between views.