Editing "a 3d image of three balls and a cube" to add concept: sunglasses

#21
by Insoo - opened
Source image:
Target image:
Target image prompt: a 3d image of three balls and a cube
add concept
sunglasses

like this a lot!

btw why does it remove two spheres in the original image and put sunglasses instead? is it the way it work as expected?

Editing Images org

So intuitively, this is what we think happens-
During the denoising process, as the model is conditioned on 'sunglasses' and doesn't start from pure random noise + has the noise maps obtained from the inversion of the input image to enforce semantics of the original image, it transforms the spheres into glasses because they resemble 'sunglasses' the most (our of the elements in this image), so it steers the denoising process in that direction as if the noisy spheres are the start of sunglasses.
This is also likely why in the attached image, all baseballs were transformed into pompoms :)
image.png

Sign up or log in to comment