GANSketching

Why we included this project

The code behind the ICCV 2021 paper on steering a pretrained GAN with hand-drawn sketches is the most direct way to see that idea in practice. If you work with StyleGAN2-style generators and want to reshape what a model produces, for example making cats stand up or altering a face's pose, without retraining from scratch, this repo shows how to fine-tune an off-the-shelf generator so shape and pose change while the rest of the image, including color and texture, stays intact. The edited model also keeps its latent space editable, so you can still apply GANSpace-style edits like adding fur or closing eyes. It ships training scripts, evaluation code, datasets, and pretrained weights, plus a Colab quickstart, which makes it practical to reproduce the results and adapt the approach to your own generator. Researchers and engineers working on controllable image generation will find a self-contained starting point here.