search

XPixelGroup / DiffBIR

Official codes of DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior
Apache License 2.0
3.33k stars 283 forks source link

Training set for BIR task #64

Open EternalEvan opened 1 year ago

EternalEvan commented 1 year ago

Hi, you choose the ImageNet for training. Did you select the images larger than 512-size before training or take all images into training set and resize them to 512-size? Thanks for your early reply!

0x3f3f3f3fun commented 1 year ago

We did not filter out those images larger than 512, but instead used the center crop method to produce 512-sized images 😄.

YilanWang commented 7 months ago

I found that in the paper, the resolution of imagenet data is set to 512x512, but as far as I know, the resolution of imagenet is 256x256. How do we handle it? Is it similar to upsampling? Thank you very much

EternalEvan commented 7 months ago

I found that in the paper, the resolution of imagenet data is set to 512x512, but as far as I know, the resolution of imagenet is 256x256. How do we handle it? Is it similar to upsampling? Thank you very much

Yeah. I assume the author upsamples the images along the shorter edge and uses the center crop method.