If anybody out there needs a good laugh. Try getting Stable Diffusion to generate a peeled banana. Good luck, and don't forget to share your best result. I'll start...
As funny and distracting as this is. It actually highlights an area where SD lacks training: peeled fruit. And really, the concept of peeling in general. It does not seem to have a true grasp of peeling layers. After a few rounds of prompts with such terms as "peeling skin" (and a little eye-bleach to go with them) I could not come up with anything at all even remotely resembling a discarded banana peel, which was the original intent.
Even framing the prompts in a way that distinctly avoided the word banana did not help. I tried every possible way I could imagine to describe the geometry of a banana peel to force an approximation.
The closest I came was using kelp as an analog. This led to a few promising results but I never quite got anything useable.
prompt = "one clump of kelp strands, bright yellow, dry on the ground. photorealistic, 8k, closeup, low angle"
seed = 866440854
guidance scale = 30
Maybe it was just the tool I was using, I don't know. If anyone has any better luck I would be interested to hear about it. I know some models do not have this issue, for example ChatGPT had no problem generating an image of a peeled banana. So maybe there's a finely-tuned fruit generating model out there somewhere that's just a well kept secret.
Top comments (0)