Open dmarx opened 11 months ago
i bet with a controlnet, we could compute strong metrics on just a handful of faces.
oh, the controlnet version might also make DINOv2 [CLS]
a more viable representation. would be nice to be able to evaluate this with a fairly "independent" representation, which CLIP definitely is not (as a component of SD, it's text encoder, i.e. the CLIP representational space is part of what SD is learning/aware of, so it makes sense that images cluster so well in it).
log size and location of bounding box, rotation angle, etc. get distributions on these statistics
compute diversity metrics on cropped/centered/rotated/aligned faces
yo. i bet i could invert idea and use a facial pose controlnet to just get ultra standardized face portraits.
that would run the risk of introducing new biases from the controlnet. could benchmark against non-controlnet methods (i.e. gptdiversity128)