This repository has been superseded by https://github.com/halcy/AnimeFaceNotebooks - probably have a look at that instead.
This repository contains scripts that can be used to calculate regions of interest for recognized tags on a given image, using the DeepDanbooru tag estimation model (keras conversion, weights can be downloaded here, distributed with permission of the models creator).
Some example outputs:
One somewhat accidental use of this is automatic censoring of images by blurring out regions that correspond to explicit / questionable tags. Code that does this can be found in this notebook (warning: includes output, which is to say, a censored explicit image).
The codes speed has been improved somewhat, but suggestions for how to speed it up more very welcome.