CartoonGAN resurrected in TensorFlow 2.0 Alpha, now archived

What it does

This repository is a ground-up implementation of CartoonGAN, the CVPR 2018 network that transforms real-world photographs into cartoon or anime imagery. It provides training scripts, inference pipelines, and two ready-to-use pretrained checkpoints that apply Shinkai and Paprika artistic styles. The project also exports trained generators to TensorFlow SavedModel and TensorFlow.js formats for browser or runtime inference.

The interesting bit

The authors grafted a lightweight generator inspired by ShuffleNet V2 onto the original architecture, giving up a little fidelity for faster inference. They also wrapped the entire research paper into a practical kit: live TensorBoard monitoring, edge-smoothing preprocessing, and a one-click web demo. It is essentially a reference implementation frozen in time.

Key highlights

• Ships both the original CartoonGAN generator and a --light variant tuned for speed.
• Includes pretrained light_paprika and light_shinkai checkpoints (~7MB) plus exported SavedModels (~11MB).
• Exports trained models to TensorFlow.js and SavedModel without extra scaffolding.
• Integrates TensorBoard to visualize losses and generated samples during training.
• Provides a browser demo and a Google Colab notebook for inference without setup.

Caveats

• Explicitly unmaintained; the authors warn that reproducing results requires following the exact environment and training instructions.
• Built against TensorFlow 2.0 Alpha, so dependency conflicts with modern TF/Keras are likely.
• Requires manual dataset curation, including a smoothed-edge variant of target cartoon images, and GPU memory limits dictate batch-size tuning.

Verdict

A solid choice for researchers or hobbyists who want a complete, reproducible CartoonGAN reference. Avoid it if you need an actively maintained library or a modern, plug-and-play style-transfer API.