search

lyh-18 / DegAE_DegradationAutoencoder

Codes for CVPR2023 paper "DegAE: A New Pretraining Paradigm for Low-level Vision"
56 stars 1 forks source link

Advice Needed for Training with Unpaired Images and Model Selection #6

Open lx123-j opened 3 weeks ago

lx123-j commented 3 weeks ago

I would like to ask how I can train to achieve better results. I used unpaired images for training, but the results were not very good. Could you advise on how to improve the training and which models should be used as DDG and DDE?

lyh-18 commented 3 weeks ago

Hi, this work is not proposed for unpaired image training. It needs paired data for downstream task finetuning. The pretraining phase is self-supervised, but it still incorporates paired images by automatic synthesis.

DDE can be a well-trained bicubic SR model, while DDG is obtained by DegAE pretraining. Our trained weights have been released and you can directly adopt them as a good initialization.

Thank you for your interest in our work.