Comment by CWuestefeld
10 days ago
I'm wondering if I read the same article. Yeah, I looked at every single one the results. And although it's been a couple of hours since, I don't recall ANY examples where it completely failed to do anything. Can you point me to an example?
I don’t want to go through every image, for the mountain:
It failed Remove background, Isolate the background, long exposed (kept people), Apply a fish-eye lens effect (geometry incorrect), Strong bokeh blur (wrong blur type)
Some were more ambiguous. Give it a metallic sheen looked cool but that isn’t a metallic sheen and IMO it just failed ukiyo-e Japanese woodblock print style but I wouldn’t object to calling it a vaguely Japanese style. Compare how colors blend with ukiyo-e woodblocks vs how OpenAI‘a sky is done.
Removing the background is impossible - or more to the point, it would yield a blank image. There is no foreground in the image, it would wind up removing everything. Which also means that its result for isolate the background is exactly right. Although we might want to argue that the lower part of the image is a midground, that's ambiguous.
You're mostly right to criticize the fisheye - it's plausibly a fisheye image, but not one derived from the original. For bokeh, you're right that it got the mountain wrong. But it did get the other samples, and it's the only one that seems to know what bokeh is at all, as the other models got none of them (other than Seadream getting the Newton right).
For the "metallic sheen", I assume you mean where they said "give the object a metallic sheen", since the first attempt had OpenAI giving the image itself a quality as if it were printed or etched on metal, arguably correct. But for that second one, for all but the 4th sample, OpenAI did it best for mountain and rubik's cube, and no worse for cats and car. Seadream wins for the Newton.
I don't have any knowledge of the Japanese styles requested, so I'm not judging those.
I've reviewed your examples, and it hasn't changed my mind.
> I don't recall ANY examples where it completely failed to do anything
> I’ve reviewed your examples, and it hasn’t changed my mind.
I think I have a better understanding of your thinking, but IMO you’re using a bar so low effectively anything qualifies. “it's the only one that seems to know what bokeh is at all, as the other models got none of them (other than Seadream getting the Newton right).” For bokeh look at the original then the enlarged images on the car. OpenAI blurs the entire image car and ground fairly uniformly, where Seedream keeps the car in focus while blurring background elements including the ground when it’s far enough back. Same deal with the cats where the original has far more distant objects in the upper right which Seedream puts out of focus while keeping the cats in focus while OpenAI blurs everything.
In my mind the other models also did quite poorly in general, but when I exclude failures I don’t judge OpenAI as the winner. IE on the kaleidoscopic task OpenAI’s girl image didn’t have radial symmetry and so it simply failed the task, Gemini’s on the other hand looks worse but qualifies as a bad approximation of the task.