To actually answer the question of how one can determine something is AI, particularly as it grows more sophisticated with anatomy: you have to train yourself to recognize artifacts.
There is no one single unifying giveaway beyond a strange sense of uncanny that you will eventually begin to recognize the better you attune yourself, and certain models have their own unique āstylesā you can begin to recognize (midjourney and stablediffusion produce very different looks, for example). There are however a few things on which one can tend to focus.
Edges: AI, as of this post, still struggles with distinct edges of objects and figures and has a tendency to blend details together. Look for hair, ribbons, and other flowy details if present. Do they fade into other details? Look at how the hair fuses with the smoke:
Edges 2: Sometimes they will also have the edges completely avoid each other, with a foreground figure slightly warping along the edges in a way that matches the background edges, like repelling magnets:
Patterns: AI, as of this post, still struggles with patterns. Filigrees, mandalas, brickwork, scales, anything that involves a high level of intricate detail tends to get blurred together. This can be a tricky one, because a lot of artists will also fudge pattern details in looser renders, but usually in a way that makes sense as an impression and notā¦. this:
Architecture: Are there buildings present in the image? AI has a tendency to make Escher-esque nonsense structures, with pillars in places they donāt belong, arches that go nowhere, bricks that donāt align, and support beams that start on one plane and connect to another. It also struggles with perspective, but, so do many humans so I would not consider it evidence alone. Check out the placement of this pillar, and also the detail on theā¦ window? Candle cage?? Thing?
Resolution and quality: AI cannot make high-resolution images. It just cannot. While most artists arenāt posting their full resolutions, generative images canāt be enhanced, and the āartistā will not be able to provide proof of work. You should be able to zoom into work by an artist and admire their strokes, relate to their errors, and appreciate their process at every skill levelā zooming into generative images somehow makes them even less clear, a mess of pixels that are somehow both blurry and also look like they have been run through a sharpen filter:
Text and signatures: AI struggles with legible characters in any language, and the result is a simlish-looking approximation of characters at worst, and hilariously misspelled words at best. Since these models are trained off real artists, they will also often have artifacts of a signature that oopsed its way into the image. These signatures are always illegible or, if ālegibleā, are not actually the names of real accounts.
Things like this can be tough to spot at a glance if youāre not actively keyed into looking for them, but theyāre the type of uncanny stuff that once you see it will start gnawing at the back of your mind. Youāll be scrolling your feed and suddenly take -1 psychic damage and you have to scroll back up to see why. Stuff that goes beyond inconsistent lightsources and bad anatomy.
gummygunk
izzy806
pixelskylines
jacobvanloon
areudirty
