← Back to context

Comment by pants2

2 days ago

At least Krea doesn't have the GPT-image shimmer! All of the GPT-image images are immediately recognizable as AI.

5 comments

pants2

Reply
jjcm  2 days ago

Or the gpt-image "fried" texture. Interestingly the best approach I've found so far is to use gpt-image for the initial design, but nano banana for the asset export once I have the design. IE in the case of the "Annui" design, I'd take that design and its backing json structure, hand it off to Opus to build it in html, then use nano banana to build out the images of the statues/background images.

You just gotta use the best parts of all of them :)

  • BoorishBears  2 days ago

    I'm unfortunately getting the gpt-image fried thing: https://imgur.com/OYY1MSa

    Which is interesting... after gpt-image-2 came out with this bug replacing the piss yellow filter, it seems like every other release except Google's have the same issue.

    Off the top of my head, this exact artifact appeared in Qwen, Grok, and now Krea.

    It makes me wonder if gpt-image-2 is being used to bootstrap post-training, or if their images are just so ubiquitous it's turned into a sort of "virus".

dvrp  2 days ago

Thank you haha

And benchmarks do not account for it sadly

Also, sometimes you want to do “ugly” things on purpose or add imperfections like film grain and such

  • nomel  2 days ago

    > Also, sometimes you want to do “ugly” things on purpose or add imperfections like film grain and such

    This is pretty trivial with an extra sentence added to your prompt, for all the models. Getting them to make normal looking people is trickier.