It seems to be a mix of WD tagger and natural language caption processing, and there's also a feature "named XXX who"...
Although most of the detections are incorrect,
it’s useful to use a script to replace them place.
However, I’ve found a problem: it’s easy for legs to be misinterpreted as "legs apart,"
for example, if the action is "crossed legs" but it gets detected as "legs apart."
I’m not sure why. If you use the original large model and WDtagger, it still detects "crossed legs."
It might be necessary to check how the dataset is constructed.
It seems to be a mix of WD tagger and natural language caption processing, and there's also a feature "named XXX who"... Although most of the detections are incorrect, it’s useful to use a script to replace them place.
However, I’ve found a problem: it’s easy for legs to be misinterpreted as "legs apart," for example, if the action is "crossed legs" but it gets detected as "legs apart."
I’m not sure why. If you use the original large model and WDtagger, it still detects "crossed legs." It might be necessary to check how the dataset is constructed.