Spaces:
Sleeping
Female figures generated by the demo
The AI does not seem to recognize when the person in the submitted photo is a female (with breasts). I've noticed that in my experiments, any female meshes that are generated do not have breasts, instead they are usually treated like males and either have very flat chests or masculine pectorals.
Perhaps this can be solved by including an option in the demo to let the user specify that the person in the photo is a female (or at least, has actual female breasts) so that the AI will handle the breasts correctly?
The 3d pose estimator is PyMAF, which uses neutral SMPL as the body template.
Github issue: https://github.com/HongwenZhang/PyMAF/issues/3
The idea is simple. We have known distributions over male and female body shape parameters *in* the gender-neutral space. If some of your training images have gender labels, then the body shape prior uses the distribution corresponding to the label.
— Michael Black (@Michael_J_Black) August 5, 2022
The way to address this is how we did it in PIXIE (https://t.co/2TWG8hmuDQ). We keep a neutral body shape space but the network learns to estimate male or female shapes within that space when it is confident about gender. This is nice because it supports non-binary genders.
— Michael Black (@Michael_J_Black) August 5, 2022