>>105883614
There is no actual 'tag', it's just captions turned into tokens which are then part of the pattern recognition mechanism of ai training.

From the ai perspective, it doesn't matter if you train with the caption style:

A blonde woman is wearing a polka dot bikini and standing on a beach, it's a sunny day.

or:

blonde, woman, polka dot bikini, standing on beach, sunny day

The training will learn how to associate each of these captioning strategies with the training images as long as you are consistent and will give equally good inference results if you prompt accordingly.

The reason we are seeing a strong shift away from 'tag' style towards 'natural language' is because the latter is the way 99% of people who are asked to write a prompt for what they want, would write it, as the name 'natural language' implies.