Look for a long thing that is briefly interrupted by something covering it. A good one is, say, the button line of a shirt that has a crossing shoulder belt or the like. Another is a long branch that goes behind another object like a person. Yet another is a walking stick that goes behind a cloak or sword or something. Almost invariably the line does not match on each side of the interrupting object. Often it doesn’t even continue at all. So with the buttons, you’ll see the line of buttons, the belt … and then just fabric. Or the buttons have jumped to the left by half a torso. Or some other such artifact.
Talking of buttons, look for buttons that don’t do anything. Like a line of buttons on what looks like a sweatshirt or the like: i.e. no gap that needs buttoning. Or buttons that look like they’re buttoning the shirt into the jacket. Or buttons that parallel the main, legit button line temporarily for no reason. (And no, I don’t mean double-breasted jackets!) Or buttons that change styles for no reason. Slopmongers have a very hard time making buttons consistent.
Text. 'Nuff said. Even the newer models still fuck this up, usually either by having gibberish that only resembles writing if you flash over it quickly with your eyes, or by having text with terrible spelling, weird artifacts that has the text growing into or out of other things, or text that sounds like it was robotically generated.
Appendages. It’s amazing to me that after all this time and money was spent, you still can’t get consistent numbers of fingers, thumbs, toes, or even primary limbs. We all know the finger thing, but how may times have you noticed the phantom extra hand or two?
Eyes. Eyes are usually well drawn, actually, but they tend to look everywhere and anywhere but where it might make sense.
All of these flaws (and many, many more) are a result of the slopmaker not having any kind of model of what it’s making. Lines stop or shift for no reason because it doesn’t know that the line on one side of an obstacle is the same as the line on the other. Eyes don’t look at sensible things because it has no idea what the picture is of and thus no idea where eyes would naturally be directed. In general look for things that require a coherent mental model to do right and you’ll spot the AI in no time flat.
All these can be touched up if the maker is actually trying to put any effort into it.
But if they’re just using diffusion as a stencil or editor, and they disclose the process… That’s fine, really. That’s more akin to extensively using photoshop.
There are many tells for AI art.
My favourites are:
Look for a long thing that is briefly interrupted by something covering it. A good one is, say, the button line of a shirt that has a crossing shoulder belt or the like. Another is a long branch that goes behind another object like a person. Yet another is a walking stick that goes behind a cloak or sword or something. Almost invariably the line does not match on each side of the interrupting object. Often it doesn’t even continue at all. So with the buttons, you’ll see the line of buttons, the belt … and then just fabric. Or the buttons have jumped to the left by half a torso. Or some other such artifact.
Talking of buttons, look for buttons that don’t do anything. Like a line of buttons on what looks like a sweatshirt or the like: i.e. no gap that needs buttoning. Or buttons that look like they’re buttoning the shirt into the jacket. Or buttons that parallel the main, legit button line temporarily for no reason. (And no, I don’t mean double-breasted jackets!) Or buttons that change styles for no reason. Slopmongers have a very hard time making buttons consistent.
Text. 'Nuff said. Even the newer models still fuck this up, usually either by having gibberish that only resembles writing if you flash over it quickly with your eyes, or by having text with terrible spelling, weird artifacts that has the text growing into or out of other things, or text that sounds like it was robotically generated.
Appendages. It’s amazing to me that after all this time and money was spent, you still can’t get consistent numbers of fingers, thumbs, toes, or even primary limbs. We all know the finger thing, but how may times have you noticed the phantom extra hand or two?
Eyes. Eyes are usually well drawn, actually, but they tend to look everywhere and anywhere but where it might make sense.
All of these flaws (and many, many more) are a result of the slopmaker not having any kind of model of what it’s making. Lines stop or shift for no reason because it doesn’t know that the line on one side of an obstacle is the same as the line on the other. Eyes don’t look at sensible things because it has no idea what the picture is of and thus no idea where eyes would naturally be directed. In general look for things that require a coherent mental model to do right and you’ll spot the AI in no time flat.
Not long before we see a lot of art that is created manually to look like AI slop, where they intentionally do that stuff.
That won’t last long as it gets dumped on the heap of stuff people don’t want to look at.
All these can be touched up if the maker is actually trying to put any effort into it.
But if they’re just using diffusion as a stencil or editor, and they disclose the process… That’s fine, really. That’s more akin to extensively using photoshop.