search

mandalinadagi / WGSR

56 stars 4 forks source link

Can I only do 4x SR? #8

Open ButoneDream opened 6 months ago

ButoneDream commented 6 months ago

When I modify the following line of code to upscale to 2, the result looks very bad. Why? model = arch.RRDB_Net(3, 3, 64, 23, gc=32, upscale=2, norm_type=None, act_type='leakyrelu', \ mode='CNA', res_scale=1, upsample_mode='upconv') like below it is bad image

mandalinadagi commented 6 months ago

We conducted experiments on x4 for WGSR, but from our other works wavelet losses works fine for x2. Did you start training x2 model from pretrained x2 RRDB?

cgysjy commented 4 months ago

Hello, I encountered the same problem while testing x2 and x3 times images using the x4 pre trained model. Have you solved it yet? Or can we only retrain x2 and x3 times the network？