A 99% automatized pipeline to construct character pack training set from anime

Overview

Project page: https://github.com/cyber-meow/anime_screenshot_pipeline
For update please see: https://civitai.com/articles/3189

Preparing data for your fine-tuned model may seem hard. It seems to become even harder for character pack, especially if you use anime sources which do not come with built-in character tags. However, anime is a good source for training for its high consistency, and with this project, all you need is to run

python automatic_pipeline.py \
    --start_stage 1 \
    --end_stage 7 \
    --src_dir /path/to/video_dir \
    --dst_dir /path/to/dataset_dir \
    --character_ref_dir /path/to/ref_image_dir \
    --image_type screenshots \
    --image_prefix my_favorite_anime \
    --log_prefix my_favorite_anime

Building a character pack dataset is now as easy as pie!

More details can be found in the GitHub repository. As I only use Linux, there can be bug if you use Windows but contributions are welcome.

This project is largely inspired by the fully automatic process of narugo1992.