[Feature]: DAAM cross-attention analysis to get an SD model's understanding of a detection prompt

[groinge]

https://github.com/castorini/daam

This has been around for a while. It could be used for semantic object recognition through img-to-img, and while obviously it lacks the precision of SAM and GD I'm thinking it could be used to trim down existing masks when SAM is incapable of masking something in the image.