Yesterday I played with inpainting Ianās image to turn it from square to portrait. Hereās my findings.
This was the starting image š¤Æ Any other AI art addicts here? š¤Æ - #236 by Kamiccola
The verdict is in, Dalle2 is great for inpainting as it can take an upscaled image and you can little by little add to it and in the end you end up with an already upscaled image.
The drawback is that if you use a large image to start with, then the inpainting area is tiny (like a tenth of the image in the case of the 3k image I was working with). So not only it takes a lot of credits to finish, it also takes very meticulous prompt. By that I mean, if AI canāt tell what the starting image is (and it canāt when itās that large), then it can make mistakes and give poor results.
So for example, I wanted to add more image above Ianās head, maybe add to the supernatural mood. His head is cut off, right? So I put in the prompt, portrait of a young man magic swirl spectral oil painting.
Result: it added a tiny man on top of his head.
Why? Because the inpainting area was so tiny that it couldnāt tell that there was already a man in the picture.
So I had to be careful with the prompt. In the end, I created something (I could post it later when I get home) but I wasnāt 100% happy and all out of credits.
So I went back to NightCafe and tried their inpainting.
Drawback is that I think it lowers the quality when you upload an upscaled image but it can be upscaled later anyway.
So with NightCafe inpainting is done by creating a mask on an existing image. Itās all done at once, not little steps like in Dalle2. So I was able to try a few times and with a few different prompts. In the end, I probably used only 5 credits or so until I was happy.
I might try again later but for now it will do.
So all of this effort was to create a new cover (in Canva).