[ticulatedspline]

https://www.aventos.dev/our-work

dunno, I've seen worse and some of the scenes if you asked me if it was AI or not I probably couldn't tell you. most egregious thing I saw was at the 27 second mark where the girl is inside the table.

At the current level of tech it feels like it would be a pretty dull full length Anime, and hard to squeeze all the bugs out but for something shorter like a commercial or FMVs for like a video game seems good enough.

[deathanatos]

0:19 - the bowl has soup in it, despite coming from the middle of the stack.

0:19 - [same time] several bowls are just sloppified, e.g., having two bottoms

0:45 - last of the soup phases out the bottom of the bowl

Continuity errors with the smock. (Is there a logo on the pocket, or no? Does it have a pocket?)

("How Asians Pay for the Bill" … what a title)

0:29 - a water drop drops off the ice cube and falls to the next ice cube … inside the water?

The cards are never quite the same; sometime the contacts are misplaced, sometimes they're missing.

But things like the girl phasing in the table, the soup phasing through the bowl … these are things a human would never animate, because they've lived their entire life understanding the physics of soup. Or, like with the card … they'd get a reference.

[qwerpy]

I watched the whole thing and didn’t notice anything off, other than soup moving a bit weirdly. But if no one had told me it was AI, then I’d think that’s just the animation style. Then I read the list of tells and oh ok, now I can see a little of it if I pause at just the right time and replay it a few times. But who watches anime like that? I think it’s good enough for the vast majority of people, who would enjoy it at first, then later find out it’s AI, and hate it because that’s the cool thing to do.

[deathanatos]

So, yes, some of this I'd've missed if I hadn't been combing over it quite as finely. But not all of it.

But in particular, the cut of the bowl stack being spilled … that scene in particular makes no sense. The entire physics of it should feel wrong because they are wrong. In the same way that CGI animations and/or games get more realistic looking the closer they simulate physics, so too does animation: if a character drops a bowl (in a setting with gravity), then it must act so, and animators know this and render it as such. Yes, and AI will try to predict it, but it strikes me akin to trying to perform a computation in screen space that must be performed in 3D space.

The soup also feels wrong, and it only takes a second rewatch to determine why; the shot feels non-committal, because it is: are we ladling, are we pouring, is the soup phasing out the bottom?

The best animators out there, to me, reward you for watching harder. Small things that make sense in the context of what you're seeing, but aren't necessary. The more you look, the more detail there is to the scene. I'd offer Frieren as an example here. For example, in E21 (the episode where Frieren breaks the barrier during the test): the episode starts when Flamme brings Frieren before Serie; this occurs hundreds, if not more than a millennia, prior to the main events. We see Serie in a temple. Later, when Frieren breaks the barrier, we see Serie remark: "This is why I will never retire."; the same temple, but now the walls are worn with time. This is made somewhat apparent, as there's a fade to denote the passage of the time. (Less apparent is there is allegedly a one or two frame shot of Serie, whose barrier it is, overlaid in the spell Frieren casts to break it. It's visible if you watch it, but not really intelligible.) As another example, the timeskip in E1 has her visiting the same shop twice, and the aging of the shop keeper & shop is visible. It's not just the same shop, it's from the same angle, specifically so that we the viewer can understand it to be the same shop. It's not required for the plot, but it builds out a richer story. As a final example, E17, Fern comes in from the cold & touches Stark on the face ("You're a monster!", Stark screams.) Later, when Fern falls ill, Stark tries to wake Frieren, who is by this point notorious for sleeping too long. Stark wakes her by touching her face, to which Frieren responds, "You're a monster." (Honestly Frieren is one of those animes that's probably so littered with these things that I'm missing like 80% of them.) But good animations make detailed choices all the way down to that level: as far as the plot cares, Stark only needs to wake Frieren, but he wakes her in the particular way that makes sense for the character we've built.

But in AI pieces, the detail is just never there, because when it's "wall in background", there's so many walls, in so many backgrounds, all muddled up in the training data.

Frieren's earrings are actually somewhat similar to the smock comment: her earrings are visually consistent throughout the entire anime, not merely a few shots. (Again, this is an example that's rich in detail, too: chronologically earlier shots of Frieren notably lack the earrings, which is made possible to notice by their consistency & distinctiveness when they're present in the first place. We're probably to presume she acquires them at some yet unrevealed point, and the visual similarity to her teacher Flamme's jewelery also leaves the viewer to wonder if they're related. But it's that choice, that these two sets of accessories appear similar, that allows us to build that connection. It's also the sort of thing that leaves me wondering "how long were the writers planning?!")

(I'd also offer the soundtrack as something that I don't know that AI can ever reach. The song "Knife to the Throat" gives me chills if I play it in my mind. But it's harder to demonstrate detail and thematic elements in a musical piece, I think. An a musical example, I'd offer "Pink Pony Club": the opening notes to which are a musical allusion to "I Will Survive". The notes are chosen intentionally, as a specific reference (one thematically relevant to the song), to set a specific tone and setting for the song that is to come, and you need to know the reference and the history and context it exists in, to be able to appreciate it.)

[qwerpy]

Thanks for the reply and the example. When I used to watch anime I always sped through it wanting to get to the story/action. I had assumed people combing through frame by frame were just looking for excuses to label it as AI. But now I can appreciate doing it for a lovingly-crafted anime to savor the details, and being disappointed when the they're not there or hallucinated/inconsistent due to AI. It's the same reason I avoid any music that is AI-generated, despite being fooled at a cursory listen. I don't want to waste my time if I dive in deeper and find out there's nothing there.

My in-laws from China are visiting and now they're both watching these AI-generated micro-dramas. Lots of weird shiny people with perfect faces and bodies, and overly dramatic yelling/crying. Much worse quality than the anime clips featured in this post. They know it's AI and don't care. I'm afraid that if the majority of people reward slop, that's what everyone's going to get.

[protocolture]

>But now I can appreciate doing it for a lovingly-crafted anime to savor the details, and being disappointed when the they're not there or hallucinated/inconsistent due to AI. It's the same reason I avoid any music that is AI-generated, despite being fooled at a cursory listen.

What gets me is that, as the amazing generation of technical animators have retired, anime has gotten less and less interesting fine detail.

AI threatens to bring that detail back, but make it wacky and crazy. And I honestly dont know which way is worse.

[hiccuphippo]

The eagle is flying in one direction and then the opposite by morphing the back into the front.

[npinsker]

One consistent tell is in hair. The most obvious manifestation is how long hair splits and merges back into itself (as at 0:27).

AI hair often doesn't follow proper "form" and looks like overcomplicated, random tufts that can clash and overlap. It might not be obvious on a first glance, but you can learn to tell the difference pretty quickly.

[protocolture]

I see people trying to make 5 - 10 minute clips of 80s anime inspired stuff, and it usually looks fine as a still, but when things move its off and when scenes change theres a vast inconsistency with previous scenes. I tend to block most of these creators even tho I should really be in the target audience.