4 ms·
> There’s also a strange preponderance of torso-only pictures—presumably the result of “fashion shots” in the training data (and, yes, with some rather wild “fa
by mk_stjames 3y ago
> There’s also a strange preponderance of torso-only pictures—presumably the result of “fashion shots” in the training data (and, yes, with some rather wild “fashion statements”):
Actually I think this is due to the staggered cropping stage on the input data- the images were all cropped to square, and if an image, like of a person, was taller than wide, the image would get cropped into a few square crops, with one of those being just the center of the image, which, for people in a tall portrait shaped photo standing up, will leave just the torso.
There was a blurb in the SDXL paper about changing the cropping preprocessing to alleviate this potential biasing.
- gwern 3y agoCorrect. Very well known artifact, you see it in all the GANs and diffusion models depending on the choice to do center vs random crop. Often changes results by several FID points (random crop adds a lot of variations and is much harder to model, confusing the NNs), which is why I suggested the cropping conditioning idea that SDXL uses with excellent results. (This is another example of how many of OP's observations are essentially minor, contingent aspects of the data or arch, and don't generalize beyond SD1.)