Hacker News new | ask | show | jobs
by SailingSperm 1388 days ago
I think it's actually the opposite. shrink the img and give it a white border, then re-imagine. Stitching together and playing the opposite way to give the zoom effect. shrink to keep same px dimensions or just enlarge canvas and place img in center.

Some good examples of this being done with Dalle2 last month -- https://youtu.be/TW2w-z0UtQU?t=244

1 comments

I don't think that's the case, it looks to be a zoom and reinterpret. I wonder what tradeoffs doing it the other way around would bring?

       rot_mat = cv2.getRotationMatrix2D(center, angle, scale) # the zoom variable is passed as scale
https://github.com/deforum/stable-diffusion/blob/5241ce95058...

Edit: having looked carefully at the video, these are pretty different. Inpaining keeps the original crop the same, whereas this version allows the model to reinterpret the original.