Experimental opinionated aesthetic score models. See GitHub for more info
CityAesthetics - Anime
Training/inference code | Live Demo
Intentional biases
- Completely negative towards real life photos (ideal score of 0%)
- Strongly Negative towards text (subtitles, memes, etc) and manga panels
- Fairly negative towards 3D and to some extent 2.5D images
- Negative towards western cartoons and stylized images (chibi, parody)
Issues
- Tends to filter male characters due to being underrepresented in the training set
- Requires at least 1 subject to be present in the image - doesn't work for scenery/landscapes
- Noticeable positive bias towards anime characters with animal ears
- Hit-or-miss with AI generated images due to style/quality not being correlated