Prompt to ChatGPT was “Draw a completely white picture”, and it sent to DALL-E:
a completely white canvas, with no features, textures, or elements, just a uniform white color
And got the above picture and this one:
Kind of interesting see what each color generates. I also tried transparency (which I imagine it can’t do anyways) with:
a completely transparent canvas, where every pixel has an alpha value of 0, resulting in a fully invisible image with no features, textures, or colors
And it gave me this:
It seems likely that example images for these kinds of prompts were removed while cleaning up the training data.
Kind of interesting when you think of how selectivity can improve and hurt a model. I think Vid2Vid stuff is necessary for most things right now because very likely training data has few people blinking, sneezing, knees bent just before jumping, and such because those make for very odd still shots.
this is really interesting… it currently cannot grasp the concept of nothing. thanks for sharing!
“white washed blank” kind of worked for me.
(https://perchance.org/ai-photo-generator)Failed :
total darkness, cosmic void, clear, , ., …
