search

csxmli2016 / DFDNet

Blind Face Restoration via Deep Multi-scale Component Dictionaries (ECCV 2020)
915 stars 215 forks source link

Why do the step 1 and 2 (crop and align face + landmarks detection) work only on cpu? #52

Open alessiapacca opened 4 years ago

alessiapacca commented 4 years ago

is there any way to make these run on gpu? what is the reason behind the fact that they only run on cpu?

FlowDownTheRiver commented 4 years ago

Check this video and comments under.So you can run it on gpu.It is possible to have gpu speed. https://www.youtube.com/watch?v=Kf9JCUuFO2o

csxmli2016 commented 4 years ago

is there any way to make these run on gpu? what is the reason behind the fact that they only run on cpu?

Hi, just focus on the performance on face restoration. I did not optimize the whole code due to the limited energy.

FlowDownTheRiver commented 4 years ago

Hi there @csxmli2016 .I don't wanna open a new thread but,been using your program for awhile now and I like it very much.So I would like to thank you.Also I am using PSFRGAN which is capable of doing multiple faces at once.I have a suggestion if you can do it.DFDNet is giving more quality restoring the faces vs PSFRGAN and PSFRGAN is able to do it on multi face,so is it possible you go collaboration each other as developers and make DFDNET work on multi faces like PSFRGAN?You can examine the code of PSFRGAN. Or in reverse order, maybe a compatibility for PSFRGAN to run DFDNet model.It would be really nice to see the developers work together.Also another one is maybe you can inject a super-resolution model in between for the background recovery.I did it in the video with a manual method with bats but it can also be run with for example realsr in between the processes.Why?Cos I have seen your objectives to achieve and I strongly believe you can do it!

JohannesTK commented 4 years ago

+1