Post

Not to be a party pooper here cause the perspective (and everything, really) looks wrong. But when a mirror reflects, its perspective will "shift" the perspective based on how angled it is relative to the ground plane of what it's reflecting slightly. Floors can also be tiled weirdly (mine is lol).

Once or twice (way before AI) I saw someone draw a bunch of improperly-converging vanishing lines on real photographs to try to claim perspective was ideological hokum when the reality is just that real life is full of things that aren't perfectly parallel

That's true. Lens distortion can also lack uniformity. Most photographic perspectives are actually "fisheye" distorted just very subtly. Moisture on the camera can also lead to such distortions. Likewise, even furniture, modern buildings and plazas are on hills or are a bit "loose" in straightness.

The main actual indicators here are the colours, and how the shapes are like an airbrushed slurry. The feet and shadows are also weird on the first one. The colours make no sense in that lighting/atmosphere and for that res it lacks the details/dithering. Everything resembles CGI lighting, too.

The third pic shows the vanishing lines of the shadows not meeting. They're not perspective / conal light lines, they're vanishing lines that vanish to the sun based on the shape, the shadows could be right here as we don't know the shape projected. The refractions are what don't make most sense.

I think what they're pointing out is that the perspective and light direction are inconsistent. You are absolutely right on the colors, details, and other bits. Side note: I don't like providing critique for AI as it gives them free pointers on what to "fix".

Yeah indeed. Just don't want people mistaking real photos for AI. Leads to witch hunts and can discredit actual news. Sometimes things look clearly AI but it's not clear why exactly in simple demonstrations x3 Plus the 2nd example is more CGI-oriented, and CGI can do reflection mistakes when baked.

The third one might have got me. The first two I could tell instantly. First one has nonsense text. The second one is very obvious from the lighting. That same AI lighting it always uses. But it's always good to learn more techniques to spot this stuff. I hate that we even have to.

It ain't a simple fix. The reason these things don't line up is part of how the system works. It's using statistics to blend a lotta sources together. But it can't tell how anything works. The training data doesn't record the angle of the light. So it's just, well, eyeballing it.

This *is* the improved version. The first round of nightmare horrors had far less consistent lighting, and way more stuff not lining up at all. Doing "better" is definitely possible, but you'd need to make the pixel-blending bot do way more work.

I think it'd have to start with a rough sketch, with layers, details, and then light sources... and now it's not just approximating the final product, it's made 3D CGI, and it took a week and more tokens than Chuck E Cheese. Like, at some point, the next stage of improvement is "invent holodeck".

Linear perspective is an illusion. Lines *appear* to converge to a single vanishing point but in reality things closer to you look bigger and things further away look smaller and that's kind of it. Linear perspective was invented (literally) to help understand the phenomenon.

I have drawn lots of perspective lines like this on photographs. The constructed lines are subject to errors, if you are extending the line from a short segment, or make an error in the point you draw through. The constructed lines don't always meet perfectly.

We will all have to be more and more cynical and questioning as time goes by, because it’s only a matter of time before they can fake these details as well.

I'm not sure this is something you could easily tailor AI to correct. It relies on very subtle properties of reality that take no effort to reproduce in a photograph, but would require a faking human serious attention to that specific problem & working a ruler. Current genAI works more holistically

You still need something that understands perspective to interact w/ the genAI at the correct software layer, and that interaction needs to work (i.e. if it corrects the perspective the image still needs to look good, & if it doesn't and you re-run it through the genAI the perspective needs to hold)

that's the horrifying job of Cybersecurity. or in this specific case Digital Forensics. which is technically just the Investigative Version. Digital Forensic Scientists tend to have a degree in Cyber Security. you have higher chances if you specialize in certain aspects like Social Engineering.

i love this in theory but i'm not looking forward to photographers using geometry-bending lenses being harrassed. it's very possible to make lenses (especially using curved mirrors) that create incoherent vanishing points

I know nothing of art or math so all I see are lines. Be ruthless to me because I know you will and tell me what I should be seeing here since the descriptions are not helping out at all. Like, I need something to compare the fakes to as far as these dotted lines are concerned.

I suspect many photos within my home would have similar vanishing point issues as are found in that hallway image due to real life construction (especially after settling) not actually being square and level! But there are plenty of other problems with that image.

Those are good, but not everyone knows to see details of guns and uniforms. I spotted a deepfake alleged historic fashion photo because I have a good knowledge of fashion and I could see why it was nonsense. But not everyone would. Perspective lines are one tool a lot of people can use to check.

It is funny how if you a minimal understanding of how weird animals appear you can spot 99% of their IA versions because clearly most IAs don't have enough references to generate a platypus or a Brazilian maned wolf. But yeah, not pratical to teach everyone about platypus anatomy =P

The same was with the visualization of hands. Young artists need a long training to draw naturally looking hands. The same problem had early AI images. But increasingly AI is improving, so now just because hands are not warped, the image might still be artificially generated.😉

Literally, hahaha. I'm always the one who has to break it to people in my family. You have to have such a broad knowledge of wildlife to be able to do it really consistently, though. Average person knows almost nothing about animals and it makes me so sad!

You're gonna find this hard to believe, but I didn't notice the chains because the helmets, the guns, and the patches were so nonsensical.

Lol, basically whatever catches your eye first is all nonsense. I don't know how people are fooled by this slop. I blame decades of underfunding arts education.

uh? Isn't the last image plain wrong? Because the light rays are practically parallel for all intents and purposes, they do not conform to a vanishing point?