L2M-GAN

Unofficial PyTorch implementation of L2M-GAN.

Steps

1. Download the wing.ckpt and put it in ./archive/models/.
2. Download the CelebA-HQ dataset from here.
3. Use the script ./bin/split_celeba.py to generate the dataset split, rename the generated folder to celeba_hq_smiling and then put it in ./archive/.
4. Make the shell script executable: chmod u+x ./scripts/train.sh
5. Execute the shell script: ./scripts/train.sh

TODOs

• [x] Implement the models.
• [x] Implement the loss functions.
• [x] Make it runnable.
• [x] Start the experiments.

Results

Experiment #1: Attribute Smiling

Final best FID: 16.93 (100k iterations)

The first row is the origin images, the second row is the smiling one and the third row is the non-smiling one.

Experiment #2: Attribute Gender

Final best FID: 32.83 (100k iterations)

The first row is the origin images, the second row is the female one and the third row is the male one.