## TODO
* Find very-high loss images based on both mainstream models and our internal ones, figure out if they're worth excluding or fixing.
	* probably need a work queue system for this so we can dispatch work to the human.
* Exclude images with too many characters tagged - they're unlikely to be classified because most of the information is erased on downscaling.
	* Example: https://danbooru.donmai.us/posts/3735555