3 ms·

An overwhelming majority of the kind of "art" people are producing with these Slop Generators is supposed to make physical sense - it just doesn't. Have a look

by mvdtnz 2y ago

An overwhelming majority of the kind of "art" people are producing with these Slop Generators is supposed to make physical sense - it just doesn't. Have a look at the dreck in the linked article. The software was not asked to produce a picture of an impossible waterfall that couldn't exist in our physical world. Everything about the prompt indicates the user was looking for a physically real scene. The picture of the girl in a meadow even used hashtags to indicate to the algorithm that it was expected to produce an instagram-like photo, but there's a seed pod floating in mid air, her dress is missing a critical strap, the lacework of the dress doesn't make sense and the direction of lighting is totally inconsistent across different parts of her figure.

ben_w 2y ago
None of which contradicts my point, that the output's flaws are things the users themselves don't notice. For example: > The software was not asked to produce a picture of an impossible waterfall that couldn't exist in our physical world. And, given my lack of knowledge of waterfall geology, I have no idea why that's impossible. > her dress is missing a critical strap Her left shoulder? I've seen real dresses where one of those slips off sideways for whatever reason. It's like the dress version of self-opening flies. If you meant something else, I missed it. I do see the seed pod; but (perhaps due to so many photographers using fill lighting or photoshop, I don't know) I see nothing wrong with the lighting. I see nothing noteworthy about the lacework either, but I wouldn't expect to as a guy whose total knowledge of things like that is "they are often intricate and frilly". There are oftentimes things I can indeed spot in GenAI images, when I make them I often find it best to make four at a time, pick the one that was closest to what I wanted, then highlight the errors and do img2img for another cycle, for a dozen repetitions, to get what I want. But most people don't look that closely. It's like this: https://xkcd.com/1015/ https://xkcd.com/1015/

by mvdtnz 2y ago

ben_w 2y ago

None of which contradicts my point, that the output's flaws are things the users themselves don't notice. For example: > The software was not asked to produce a picture of an impossible waterfall that couldn't exist in our physical world. And, given my lack of knowledge of waterfall geology, I have no idea why that's impossible. > her dress is missing a critical strap Her left shoulder? I've seen real dresses where one of those slips off sideways for whatever reason. It's like the dress version of self-opening flies. If you meant something else, I missed it. I do see the seed pod; but (perhaps due to so many photographers using fill lighting or photoshop, I don't know) I see nothing wrong with the lighting. I see nothing noteworthy about the lacework either, but I wouldn't expect to as a guy whose total knowledge of things like that is "they are often intricate and frilly". There are oftentimes things I can indeed spot in GenAI images, when I make them I often find it best to make four at a time, pick the one that was closest to what I wanted, then highlight the errors and do img2img for another cycle, for a dozen repetitions, to get what I want. But most people don't look that closely. It's like this: https://xkcd.com/1015/ https://xkcd.com/1015/