[daveguy]

The positioning seems very consistent, almost to the point where I wonder if that was part of the selection process to demonstrate the differences in style. There are only four per style, where the position of a subject could be a selection factor. Hard to tell if the position similarities are driven by the Stable Diffusion model or by the selection of representative images.

[tough]

The composition and positioning come from the original seed. If the same seed is used, the same background image noise is applied for the initial image which is transformed into all the styles.

Thus the similarities you see would make sense if using also the same seed for the tests.

[totetsu]

Yes, if you generate a bunch of images of waves or the sea, or other repeating patterns with the same seed, you can see how all the 'peeks and toughs' of those patterns line up in the same place.

[olejorgenb]

Maybe it's based on img2img? Some are different enough that it's not obvious that's the way it has been done though.

[adityashankar]

Hey, I'm the developer of this project, no this is not based on img2img all the images just have the same seed

[michael-ax]

incredible job! i was just showing my own experiments with sd to a friend and this just take the cake. thank you so much for mixing in that artist list!!