search

royorel / StyleSDF

Other
536 stars 50 forks source link

64x64 FID of FFHQ #31

Open zkcys001 opened 1 year ago

zkcys001 commented 1 year ago

That is nice work! I have some concerns about the FID. When we use the training code of the first stage (i.e., training the volume renderer), what is the FID between the 64x64 real images and generated images? I use the gan metric code to get about 20 fid on the 64x64 resolution. This result is correct? Looking forward to your reply!

royorel commented 1 year ago

Hi @zkcys001,

We did not evaluate the FID score for the volume renderer 64x64 outputs in our work, so unfortunately, I can't give you any reference point for the quality.

zkcys001 commented 1 year ago

Hello, thanks for your reply! I implement this method in the new codebase, and hope that have the accurate result~ Another small issue is that the 'new_train.py' should be 'train_full_pipeline.py' in the script/train_afhq_full_pipeline_512x512.sh.

Best wishes~