I need a dataset for face generation using autoencoders/GAN's.
The problem with datasets (such as the FFHQ, CelebA, etc) which use cropped and aligned in the wild face images is that these have significant variation in features such as expression, direction in which the face is pointing, framing, lighting, background, accessories on the face, etc.
These features are not relevant to facial generation (atleast the way in which I want my model to be trained). What I'm looking for is something similar to the approach used in this video. The author uses his high school's yearbook in which there is little variation in the irrelevent features I mentioned before as everyone looks at the camera, in front of a neutral background, with the same pose.
However, I can't access the author's yearbook (for obvious reasons) and I would suspect the number of images would be too little anyway.
I am looking for a few thousand images, with non face-related features controlled. I suspect no dataset satisfying all these conditions would exist, but I'm hoping I can find the next best thing. The closest dataset I found for my application was the US Adult faces dataset, which has about 10,000 unique images of individuals, with an oval cut-out blocking most of the background.