Photo-to-anime model trained on actual film frames

What it does AnimeGANv2 converts landscape photographs and video frames into anime-style imagery using a lightweight generative adversarial network. It offers three distinct visual styles based on the work of Hayao Miyazaki, Makoto Shinkai, and Satoshi Kon, trained on curated high-definition frames from their respective films. The generator weighs in at 8.17 MB, and the project includes a smaller “lite” variant for resource-constrained environments.

The interesting bit Rather than learning a generic anime look, the model ingests the specific color grading, edge work, and texture density of The Wind Rises, Your Name, and Paprika, so the output carries the actual visual fingerprint of those directors instead of a one-size-fits-all cel-shader.

Key highlights

• Three style checkpoints trained on 1,752, 1,445, and 1,284 frames sourced from 1080p and Blu-ray releases
• Generator reduced to 8.17 MB with an even smaller lite version available
• Explicitly designed to suppress the high-frequency artifacts and training instability reported in AnimeGANv1
• Supports both single-image inference and full video conversion pipelines
• Colab notebooks and a community PyTorch port lower the barrier to experimentation

Caveats

• Built on TensorFlow 1.15.0 with CUDA 10.0 dependencies, making it a legacy stack by current standards
• Non-commercial license only; commercial applications require direct written authorization from the author
• AnimeGANv3 has already been released, leaving this repository in a maintenance-only state

Verdict A solid choice for researchers or hobbyists who want stylistically specific, lightweight photo-to-anime conversion. Look elsewhere if you need a modern framework or a permissive commercial license.