Does it really work like that? I would say that they are not trying to fool any test, just getting harder to be detected. The goal being looking completely realistic.
Looking completely realistic and being able to discern between real and fake are competing goals. If you can discern the difference, then it does not look completely realistic.
Some are very easy to spot, like a shadow of a character, that’s missing a limb on the shadow, or has different placement or pose. Illumination or parallel surfaces where they vary in shadowing without a reason, is also a dead giveaway. But the mist damning evidence is having one scene, then a slightly different scene in a reflection.
There are reasons for human authors to do any of these on purpose, but unless that purpose is part of the work, they’re most likely AI mistakes.
Of course it’s kind of funny how there is already a large overlap between the best AI art, and the most senseless “modern art”.
It’s quite easy to trick people with untrained eyes… for one, they have no idea what “consistent illumination” and stuff means. And something being off doesn’t mean that an AI made that mistake because humans make mistakes, too – photographs don’t, but the general problem is not just about telling realistic stuff apart but also illustrations. You’re looking specifically for mistakes that AI is likely to make, but humans are practically never going to make. And yes humans get hands wrong all the time.
Yes, my comment applied more to photorealistic AI images.
Illustrations are a different beast, where people have much more creative freedom… and that video is reasonably good at explaining that, but I find it falls short at some points:
AI image generators don’t “consult” source images to generate an output. At training time, they extract patterns from the training set, which is never again used for generation, only the extracted patterns are.
Modern AI generators are increasingly good at generating text. They still struggle a bit, but compared to a year ago, they can now generate headlines and large text correctly, while the mess gets shoved into smaller and less important text. This isn’t all that different from human artists adding “filler gibberish” text.
Layers. While a naive (and cheaper) approach to AI generation doesn’t use layers, there are generators which do use layers, and can keep object consistency across obscured or cut-off sections.
As AI generators advance, all these differences are likely to disappear… by following this same criticisms to fix things.
AI image generators don’t “consult” source images to generate an output.
Well, you have an artist breaking things down for an audience understanding neither the technical nor artistic aspect…
Modern AI generators are increasingly good at generating text. They still struggle a bit
I mean… SDXL still struggles a lot. The only thing you can get it to spell reliably is probably “Hooters”. There’s the one or other lora which makes it not suck completely but it’s still nowhere near actually good at generating text, the training just isn’t there. And even with that in place things like signatures are probably going to be gibberish.
While a naive (and cheaper) approach to AI generation doesn’t use layers, there are generators which do use layers,
Unless you start off training by feeding the model 3d data (say, voxels) alongside 2d projections I don’t think it’s ever going to develop a proper understanding of these kinds of things. Or, differently put: Learning object permanence (of sorts, related) is a meta-cognitive abstraction step that just won’t happen with the type of topologies we know how to engineer. It’s probably like 90% on the way towards AGI, so to get a simple topology to understand it we have to spoon-feed it permanence information alongside the (apparent) non-permanence.
breaking things down for an audience understanding neither the technical nor artistic aspect…
Not a reason to misrepresent things. Reminds me of the animistic fallacy, if they even understand what’s really going on themselves.
As for text, I’ve seen the MS generator spit out decent text, at least in titles and logos, and some AI art with full legible sentences.
Unless you start off training by feeding the model 3d data (say, voxels) alongside 2d projections
Some time ago already, there was an SD fork with bounded box support, and a ChatGPT preprocessor prompt template to do the layout. Object permanence in this case is as simple as continuing with the lower layer once the upper one is finished, maintaining object continuity in the lower layer. It’s reasonable to expect this to go from bounded boxes, to freehand layers for each object. Since an LLM has been shown to be a good preprocessor to set the layout, some more integration between both, with object feedback from the SD to reduce the layer bounding box, would do wonders. Adding an opacity mask could be a bit harder, but sounds doable.
I don’t see the need of much higher abstraction to address this issue. Rendering videos of translucent objects, might need it, though.
Another arm in the arms race. The next gen of face generation will have this mastered.
Isn’t it a face race, then?
Does it really work like that? I would say that they are not trying to fool any test, just getting harder to be detected. The goal being looking completely realistic.
Looking completely realistic and being able to discern between real and fake are competing goals. If you can discern the difference, then it does not look completely realistic.
I think what they’re alluding to is generative adversarial networks https://en.m.wikipedia.org/wiki/Generative_adversarial_network where creating a better discriminator that can detect a good image from bad is how you get a better image.
This is one of the basic techniques to spot AI fakes:
The “test” they’re trying to fool, is kind of the Turing test: whether humans can tell them apart.
Consistent illumination and shadows is a rabbit hole we really don’t want to hop into.
Outside of very obvious anomalies even a trained eye will have a hard time discerning what’s going on.
Some are very easy to spot, like a shadow of a character, that’s missing a limb on the shadow, or has different placement or pose. Illumination or parallel surfaces where they vary in shadowing without a reason, is also a dead giveaway. But the mist damning evidence is having one scene, then a slightly different scene in a reflection.
There are reasons for human authors to do any of these on purpose, but unless that purpose is part of the work, they’re most likely AI mistakes.
Of course it’s kind of funny how there is already a large overlap between the best AI art, and the most senseless “modern art”.
It’s quite easy to trick people with untrained eyes… for one, they have no idea what “consistent illumination” and stuff means. And something being off doesn’t mean that an AI made that mistake because humans make mistakes, too – photographs don’t, but the general problem is not just about telling realistic stuff apart but also illustrations. You’re looking specifically for mistakes that AI is likely to make, but humans are practically never going to make. And yes humans get hands wrong all the time.
Here’s a good video about what to look for and what not.
Yes, my comment applied more to photorealistic AI images.
Illustrations are a different beast, where people have much more creative freedom… and that video is reasonably good at explaining that, but I find it falls short at some points:
As AI generators advance, all these differences are likely to disappear… by following this same criticisms to fix things.
Well, you have an artist breaking things down for an audience understanding neither the technical nor artistic aspect…
I mean… SDXL still struggles a lot. The only thing you can get it to spell reliably is probably “Hooters”. There’s the one or other lora which makes it not suck completely but it’s still nowhere near actually good at generating text, the training just isn’t there. And even with that in place things like signatures are probably going to be gibberish.
Unless you start off training by feeding the model 3d data (say, voxels) alongside 2d projections I don’t think it’s ever going to develop a proper understanding of these kinds of things. Or, differently put: Learning object permanence (of sorts, related) is a meta-cognitive abstraction step that just won’t happen with the type of topologies we know how to engineer. It’s probably like 90% on the way towards AGI, so to get a simple topology to understand it we have to spoon-feed it permanence information alongside the (apparent) non-permanence.
Not a reason to misrepresent things. Reminds me of the animistic fallacy, if they even understand what’s really going on themselves.
As for text, I’ve seen the MS generator spit out decent text, at least in titles and logos, and some AI art with full legible sentences.
Some time ago already, there was an SD fork with bounded box support, and a ChatGPT preprocessor prompt template to do the layout. Object permanence in this case is as simple as continuing with the lower layer once the upper one is finished, maintaining object continuity in the lower layer. It’s reasonable to expect this to go from bounded boxes, to freehand layers for each object. Since an LLM has been shown to be a good preprocessor to set the layout, some more integration between both, with object feedback from the SD to reduce the layer bounding box, would do wonders. Adding an opacity mask could be a bit harder, but sounds doable.
I don’t see the need of much higher abstraction to address this issue. Rendering videos of translucent objects, might need it, though.