tachibanayoshino/animegan

AnimeGAN is a Python project that converts ordinary real-world photos into anime-style images using an AI technique called a generative adversarial network. You give it a photograph, and it outputs a version of that image that looks like it was drawn in the visual style of Japanese animated films, particularly the style associated with director Hayao Miyazaki. It can also process video files frame by frame to produce animated anime-style clips. The project is a TensorFlow implementation based on an academic paper titled "AnimeGAN: a novel lightweight GAN for photo animation." The network was trained on landscape photographs paired with frames extracted from anime films. The README notes that the original training set consists mostly of landscape photos, so users who want better results with portraits of people are advised to add at least 3,000 people photos to the training set and retrain the model themselves. Using the pretrained model is straightforward from the command line: one script handles still images and another handles video files. Training your own model involves downloading a base neural network, downloading the provided dataset, running a smoothing step on the training images, then running the training script. Pretrained weights are available as downloads from the repository's releases page. Two improved follow-up versions exist in separate repositories: AnimeGANv2, which reduces image artifacts and uses higher-quality style data from Blu-ray film sources, and AnimeGANv3. A community-contributed browser-based version also exists, letting users try photo animation without installing any software. The code is free for academic and non-commercial use. Commercial use requires contacting the authors for an authorization letter. It runs on Python 3.7 with TensorFlow 1.15 and a compatible NVIDIA GPU.