AnimeGANv2 breaks up with TensorFlow

What it does

This is a PyTorch rewrite of AnimeGANv2, the GAN that turns photos into anime-style images. The author converted the original TensorFlow weights into PyTorch checkpoints, letting you run inference without resurrecting TF 1.x. It also ships a few extra distilled models tuned specifically for human faces at 256×256 and 512×512.

The interesting bit

The real value is the weight-conversion plumbing and the extra checkpoints. The README openly admits a slight fidelity loss from a PyTorch bilinear upsampling quirk, but the project compensates by offering torch.hub one-liners and a face2paint utility that handles the preprocessing boilerplate.

Key highlights

• Ships converted weights for four original styles (celeba_distill, face_paint_512_v1, face_paint_512_v2, paprika) plus two custom face variants.
• Loads directly via torch.hub, so you can treat it like a standard pretrained model.
• Includes a face2paint helper to abstract away image resizing and tensor formatting.
• Community integrations include a Hugging Face Spaces demo and a Colab notebook.
• Face Portrait v2 explicitly trades some beautification for better robustness compared to v1.

Caveats

• Converted weights produce slightly different results from the original TensorFlow implementation due to a known PyTorch bilinear upsampling issue.
• The README is inference-focused; it offers no training details or architecture notes beyond the original paper.
• Additional face weights are hosted on Google Drive rather than release artifacts.

Verdict

Worth a look if you want anime-style face transfer in a PyTorch-native stack and can tolerate minor output drift. Skip it if you need pixel-perfect reproduction of the original paper or are hoping to train your own styles from scratch.