what's the em-dash moment for image generation? I can see a whytryai leaderboard that distills it down. When i go to AI to draw me a picture, unless I include in the prompt 'make it look like my granny took a crappy snapshot,' I want dramatic.
Technically, this kind of test isn't particularly representative, since if people want a specific "look," it's quite trivial to add it in. For instance, GPT-4o defaults to oversaturated polished yellow, but it's absolutely possible to coax a rough and unpolished image out of it by adding stuff like "iphone photo," "polaroid photo," "ordinary," "amateur," "unfocused," or whever specific "crappy" look you have out of it.
I think you know how much I presently loathe GPT-beige, the grossest and most abundant internet color of all of 2025.
I do think these edge cases are going to be where a lot of the action is - even if they don't really solve it, the other models are now addressing this idea too. Interesting to see ideas getting out into the ecosystem so quickly!
Right? I think Krea 1 and Seedream 3.0 really shine in that one. One might complain about Krea man's somewhat questionable ear ridges, but I'm only looking for those because I know it's AI.
Yup, hence my "at first glance" caveat. I'm mainly looking at the style of the photo, but the composition isn't great. And this is the best of the bunch, too. Two of the other images literally had cars heading towards each other at full speed. Not the best way to stay alive on the road!
what's the em-dash moment for image generation? I can see a whytryai leaderboard that distills it down. When i go to AI to draw me a picture, unless I include in the prompt 'make it look like my granny took a crappy snapshot,' I want dramatic.
Technically, this kind of test isn't particularly representative, since if people want a specific "look," it's quite trivial to add it in. For instance, GPT-4o defaults to oversaturated polished yellow, but it's absolutely possible to coax a rough and unpolished image out of it by adding stuff like "iphone photo," "polaroid photo," "ordinary," "amateur," "unfocused," or whever specific "crappy" look you have out of it.
I think you know how much I presently loathe GPT-beige, the grossest and most abundant internet color of all of 2025.
I do think these edge cases are going to be where a lot of the action is - even if they don't really solve it, the other models are now addressing this idea too. Interesting to see ideas getting out into the ecosystem so quickly!
Can you Ghiblify this comment please?
I think the prompt 5 old man is possibly the least "AI look" image I've seen. You simply can't tell that wasn't pulled from grannie's photo-album.
Good real world tests throughout.
Right? I think Krea 1 and Seedream 3.0 really shine in that one. One might complain about Krea man's somewhat questionable ear ridges, but I'm only looking for those because I know it's AI.
Krea 1's image for prompt #4 is a bit akward...look at the car at the right, is it backwards with white lights???
Yup, hence my "at first glance" caveat. I'm mainly looking at the style of the photo, but the composition isn't great. And this is the best of the bunch, too. Two of the other images literally had cars heading towards each other at full speed. Not the best way to stay alive on the road!