Skip to content

racinmat/ahegao

Repository files navigation

Ahegao

This repository is for pix2pix style transfer between normal facial expression and ahegao.

The training set of aligned pairs is automatically created by mining ahegao videos from tiktok.

Getting videos

The python api client is used for download from tiktok. Login identifiers must be reverse-engineered from tiktok, by capturing packets. Specifically: device_id, iid, openudid. If registered to tiktok without password, password needs to be set for successful registration via API.

Listing of videos by tags works when we have hashtag ID by method list_hashtag.

For textual hashtag, we can obtain its ID by searching hashtags using method search_hashtag.

Hashtag IDs are autoincrement, thus we can crawl all hashtags over ids.

The ahegao frame in video is detected by face emotion detector.

Obtaining downloaded videos

Around 5k of videos was uploaded to uloz.to and can be downloaded from these links as 6part zip that need to be unzipped at once.

Hopefully more work will be added later.

About

Repo for ahegao detection and style transfer

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published