After one of my tweet went semi-viral, I motivated myself to put together the results of my stylegan2-ADA experiment in this post. Addition of augmentation in the discriminator of stylegan2 proved to be the cherry on the cake if one wants to train stylegan2 for a limited number of images.

In the StyleganADA paper authors explore small datasets with around 1500-2000 images. I took up the task of finding the lower limit and train on a few hundred images.

I trained on three categories namely teddy bear faces, AmongUs characters and Lego images. I used the original stylegan2-ADA repo (TensorFlow version) for training. Main points that make the training successful without mode collapse are

1. Right Initial point Fine-tuning from weight file which is close to your data helps a lot
2. Augmentation Strength If your dataset contains a few hundred images it is better to start with high augment strength and keep target low to speedily ramp up the augment strength.
3. Learning rate and Gamma Its better to explore few values Learning rate and Gamma before starting full training. Also, it is recommended to try different values when mode collapse occurs.

Below are details of a dataset, weight file link and results of each category. Hope you have a fun time watching the results!

Teddy Bear face

• Dataset: Teddy bear dataset
• No of Images: 637
• pkl file link: google drive

AmongUS Characters

• Dataset: Scrapped from citypng.com
• No of Images: 203
• pkl file link: google drive

Turn on audio 🔊🎧 for best experience. Above video is generated from Culture Shock with modifications. Music is from this youtube video.

Lego images

• Dataset: Scrapped from brickset.com
• No of Images: 31
• pkl file link: google drive

Failure case: As the number of images are very low and diversity among them is very high, despite trying above mention steps it failed to converge properly.

• 🐦Thank you for reading, if you want to comment or have any suggestions feel free to ping me on twitter.