JingyunLiang / HCFlow

Official PyTorch code for Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling (HCFlow, ICCV2021)

Home Page:https://arxiv.org/abs/2108.05301

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Does the resolution of LQ images will affect the quality of results? I use the 256x256 LQ images for inference and the results are bad

PengQiTu opened this issue · comments

No, the LQ image resolution doesn't have much impact. What kind of LQ images do you use? HCFlow is trained on (bicubic LR, HR) pairs, so it can only deal with bicubic LR images.

Thanks for your reply! The LQ image I used are degraded with gaussian noise, gaussian blur, motion blur and jpeg etc. If I use my data to train the model, can I get good results?

Yes, you can. It is just a domain mismatch problem, because HCFlow (trained on bicubic SR imaegs) has never seen the Gaussian blur and noise before. As shown in BSRGAN, ICCV2021, deep learning models have the potential to deal with complex degradation if we train them with a complex degradation model.