Diffusion-based image generators are betraying their seeds

In my article about benchmarking image generation AI I generated all images using the same seed and I noticed something:

Did you notice it, too? These images are the averages of the images generated in the benchmark article (each image here corresponds to all images generated by one image generation model). They look decidedly non-random, despite resulting from 100 different prompts averaged together. You can clearly see some structure in those image. It is even more obvious if you first run canny edge detection on the generated image and then average them (and use autocontrast to make the pattern more visible):

These patterns appear, because all images started from the same initial random image. The goal of the image generators was to just generate images, not to hide information from the original random image, and so patterns from the original random image remain visible and are in some cases (like edges) amplified.

The patterns are not the result of the used prompts. If I average images from different models (one image per prompt), I get the following two results:

The patterns are much less pronounced, because different models use the initial noise in different ways. If I choose 100 images not generated by AI, patterns essentially disappear, which is what you would expect:

And to show you that the patterns are definitely caused by using the same initial random image for all 100 prompts, I repeated the experiment using 100 random seeds, i.e. a different initial random image for each prompt. If I average those, the patterns also disappear:

I'm almost certain that fact can be used to help distinguish AI generated images from real images. The number of random seeds is finite and many are reused. Because so much structure from the initial random image is retained, recognizing that structure in the final output is likely possible and would be enough to decide that the image is AI generated. Of course there is an easy countermeasure: Just increase the number of random seeds so that reuse of the same seed becomes essentially impossible (without reuse, finding the correct seed by testing them all takes too long). But it is also possible that information from the initial random seed is retained in a way that depends only on the process, not the noise image itself, which would make recognition independent of the seed. In that case, increasing the number of seeds wouldn't work. Given that recognition of fakes will become increasingly important, this is probably an important area of research.

Paralinguistic/connotation key:

Mocking
Sarcasm, e.g. "Homeopathy fans are a really well-educated bunch"
Statement not to be taken literally, e.g. "There is a trillion reasons not to go there"
Non-serious/joking statement, e.g. "I'm a meat popsicle"
Personal opinion, e.g. "I think Alex Jones is an asshole"
Personal taste, e.g. "I like Star Trek"
If I remember correctly
Hypothesis/hypothetical speech, e.g. "Assuming homo oeconomicus, advertisement doesn't work"
Unsure, e.g. "The universe might be infinite"
2 or more synonyms (i.e. not alternatives), e.g. "aubergine or eggplant"
2 or more alternatives (i.e. not synonyms), e.g. "left or right"
A proper name, e.g. "Rome"

One always hopes that these wouldn't be necessary, but in the interest of avoiding ambiguity and aiding non-native English speakers, here they are. And to be clear: These are not guesses or suggestions, but rather definite statements made by the author. For example, if you think a certain expression would not usually be taken as a joke, but the author marks it as a joke, the expression shall be understood as a joke, i.e. the paralinguistic/connotation key takes precedence over the literal text. Any disagreement about the correct/incorrect usage of the expression may be ascribed to a lack of education and/or lack of tact on the part of the author if it pleases you.