First of all, they did do a bad job with their blog post.
…But I’m also shocked by the number of people who think they’re human AI detectors.
I suspect people are a year or two behind, looking for signs of messy Stable Diffusion XL output, like suspiciously styled appendages or that weird framing 1.5 always did. Or maybe the Ghibli-esque style Tweeters used to ape Sam Altman.
Models aren’t like that any more.
The other day, I was sitting at a TV, looking at photos side by side, I cannot tell my own mirrorless camera RAWs apart from some locally-run generations that used my photos as a reference, even if I zoom in to pixel peep or try to nitpick the depth-of-field from my lens.
Videos edited by H3 are shockingly good now, albeit at lower resolution.
I did some animesque character design mockups (just for my own thinking/brainstorming), and I can feed the model separate reference images for characters and art styles and poses and it… just gets it. It looks just like the original painted style, and I can’t find any artifacts or distortions that jump out. Not counting the ones in the original material.
I’m not trying to glaze diffusion or anything; quite the opposite. It’s messy, and sloppy. Besides, that’s not the point, and I don’t want to get into that.
What I’m saying, outside of really lazy slop or bad models like ChatGPT, people are behind if they think they can spot AI-generated stuff reliably.
So, maybe the studio lied.
The blog post certainly makes it suspicious. They could have uploaded some asset at least?
But, as suspicious as the style is, I can’t tell if that “process” picture is AI generated. Certainly not because the 2nd frame has a cartoon style, or the 3rd and 4th look like weird cg.