Encoding-Driven GAN Image Synthesis

Generate images using BigGAN that maximally drive a neural encoding model of your choice.

Examples below use a face-selective area (FFA) encoding model:

Maximized

Minimized

'

Requirements

pip install torch torchvision pytorch-pretrained-biggan matplotlib numpy tqdm pillow openai-clip scipy

Usage

Open encoder-driven-gan.ipynb and plug in your encoding model.

We have provided example FFA weights trained on ResNet50-CLIP (subject S1, Murty185) in ./weights/ so you can try it out right away.

Credits

BigGAN synthesis code adapted from Alex Abate.
Encoding model interface by Alish Dipani. Example weights from Murty185 Dataset.