That might be true, but after experimenting with DALL·E 2 last week (and spending more than $15), I have a different theory.
My tests focused on how well it could create art works around three common themes: still life, landscape, and portrait. For the first two categories, almost all the results were works that would not have looked out of place in a museum or art gallery. In contrast, with the prompt of “A painting of a young woman sitting in a chair” and variations, while DALL·E 2 produced convincing clothing, furniture, background, etc., the faces were mostly horrible. I started adding “from the rear” and “turned to the side” to the prompt just to get the face out of the picture.
I came to suspect that DALL·E 2 is bad at faces not because the developers made it that way but because human beings are uniquely hardwired to recognize faces. Most people are able to recognize and remember hundreds of faces, and we are very sensitive to minor changes in their configurations (i.e., facial expressions). When we look at a painting of a person sitting in a chair, we don’t care if aspects of the chair, the person’s clothing, etc. are not precisely accurate; a slight distortion of the face, however, can ruin the entire work. DALL·E 2 does not seem to have been trained to have the same sensitivity to faces that humans have.
If anyone is interested, the works that DALL·E 2 created for me are at [1]; video slideshows with musical accompaniment are at [2].
[1] http://www.gally.net/temp/dalleimages/index.html
[2] https://www.youtube.com/playlist?list=PLj4urky_8icRPzgFS_b98...