Unrelated to the topic, but
led by a strictly limited physical collector’s release created with game preservation at its heart
This sentence is so funny. How is a limited edition made for people who want to pay extra and have shiny object on their shelves working towards game preservation in any form? It’s so funny when they try to play “here, have more disposable plastic and pay 50 extra bucks for it! I swear you’re ALSO doing it for a good cause!”
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.
I’m not going to claim to be good at detecting AI generated content, but for visual content there’s clear vibes that AI generated content give that I can never specifically point out, but when I spot them there’s usually at least one other person calling out the slop. Yes 3-4 years ago diffusion models would make errors like mangling hands and mushy details, but while newer models get images technically correct, they leave clear vibes that are easy to spot
Motion errors are also still very much spottable on AI video. Shit blurs, but not in the way caused by actual motion being captured on film and causes distortion of the thing moving in ways that don’t happen with motion blur (like a waving arm bending at points anywhere other than the wrist and elbow).
