search

yinzhicun / MetaF2N

MetaF2N: Blind Image Super-Resolution by Learning Efficient Model Adaptation from Faces (ICCV 2023)
55 stars 0 forks source link

Model performance issues #9

Open yanglibo0512 opened 2 months ago

yanglibo0512 commented 2 months ago

When running test.py to test a picture of Celeb_iid, such as 000990.jpg, the effect is very poor. faca_hq is right. WechatIMG15 but final result is poor WechatIMG16 The parameters for running test.py are the same as default

yinzhicun commented 2 months ago

Since that the face_hq is used as the GT to finetune the model, the final result won't be better than face_hq. In order to get a better restoration result, you can increase the fine-tuning steps during inference or try to train a new model with more inner loop fine-tuning steps.