Training Data
There are many times with the malformed legs SD produces that feet are not recognisable, even at lower thresholds. I wonder why you didn't use existing data that's already labeled, from models where this works ok?
For example you can find here data (especially filtering for models with more images that have better detection results):
https://universe.roboflow.com/ashhar-ali-cnkpk/foot-detection-final-2
See tabs on the left for the data you can download as zip, etc. Plus maybe even combine data from more of those models - in case they didn't do it already.
I simply wasn't aware of existing data.
Realistically, this model was created to take care of the sort of images I normally generate. Version 1 can be very inaccurate at times, since the training folders were mislabeled. Version 2 greatly improves on version 1, but it's by no means perfect.
As I find myself training LoRAs and checkpoints these days, I really haven't gotten around to making a version 3... yet.
I only got 8GB of VRAM at my disposal, so things are slow on my end.