What it does

BlendGAN generates stylized human faces from either random latent codes or real face photos, applying arbitrary artistic styles without needing a separate model trained for each style. It ships with pretrained models, a Jupyter demo, and scripts for image generation, style transfer, and interpolation videos.

The interesting bit

The trick is a weighted blending module that implicitly mixes face and style representations inside the generator, plus a self-supervised style encoder trained on a generic artistic dataset. This avoids the classic layer-swapping approach that demands hundreds of style-consistent images per style. The authors also released AAHQ, a large-scale artistic face dataset, to support this.

Key highlights

• Single unified model handles arbitrary styles; no per-style retraining
• Supports both latent-guided and reference-guided generation
• Pretrained models and inference scripts ready for 1024px output
• Jupyter notebook demo and Gradio web demo on Hugging Face Spaces
• Built on familiar components: StyleGAN2, pSp encoder, IR-SE50 backbone

Caveats

• The README notes deformation artifacts in generated images, with a suggested workaround (add_weight_index=7)
• CUDA kernels carry Nvidia’s non-commercial license; test images are non-commercial (CC BY-NC-SA 4.0)
• Code borrows heavily from rosinality’s StyleGAN2 reimplementation; not a ground-up rewrite

Verdict

Researchers and hobbyists working on controllable face stylization should grab this. If you need commercial use or non-face domains, the license and scope will pinch.