Face Synthesis and Super Resolution Using Stacked Generative Adversarial Network
- a.Using HD CelebA Cropper to generate the dataset raw data.
- b.Use the default 0.7 face_factor.
- c.Copy the raw data to dataset folder.
- c.Run 1.get_celeb_train.py to generate training, validating and testing data
- a.Run the training command
python 2.train_and_test.py --phase train --dataset_name yourData --origin_size 64 --image_size 256
- b.Check the sample folder during training to make sure the result is correct.
- a.Run the testing command
python 2.train_and_test.py --phase test --dataset_name yourData --origin_size 64 --image_size 256
- b.Check the generated test folder for the results
- a.The FID calculating code is from here
- b.Run 3.split_testset.py to seperate blur tesing images and ground truth
- c.Run the FID calculating command
python 4.fid_score.py dataset/yourData/gt test
Table 1 FID score compare in different resolution transformation
name | 16 to 64 | 32 to 128 | 64 to 256 |
---|---|---|---|
Wavelet SRNet | 45.952 | 54.533 | 31.929 |
Cycle GAN | 63.837 | 32.228 | 42.218 |
SRN-Deblur | 125.576 | 41.463 | 12.729 |
ours | 51.646 | 29.685 | 13.431 |