When Irish filmmaker Ruairi Robinson began uploading a series of short clips created with Seedance 2.0 — TikTok developer ByteDance’s newest video generation model — it was hard to deny that the footage was much more impressive than what we’ve seen from other gen AI outfits. The clips’ star (a digital duplicate of Tom Cruise) looked a lot like the real thing as it fought Brad Pitt, humanoid robots, and zombies. And the characters moved with a complex fluidity almost passing for choreography and amplified by kinetic “camerawork.”
Gen AI enthusiasts love to proclaim that the traditionally produced entertainment industry is cooked, and some of Hollywood’s biggest studios did seem to be alarmed by Seedance’s recent capabilities as the ersatz-Cruise videos continued to rack up views online. Motion Picture Association, Disney, Paramount, and Netflix each sent ByteDance cease and desist letters over claims of copyright infringement. And in response, ByteDance said that it would take steps “to strengthen current safeguards as we work to prevent the unauthorized use of intellectual property and likeness by users.” ByteDance has yet to officially release a version of Seedance that prevents users from generating footage the company doesn’t have the rights to create.
Everything about Seedance 2.0’s rollout has felt like a viral stunt, especially when studios have already made it clear that they’re willing to sue when AI companies steal their IP. It’s true that Seedance-created videos look far better than much of the stuff we’ve seen made with Sora, Veo, Runway, and others. But the fact that churning out very polished ripoffs is the new model’s main claim to fame makes Seedance 2.0 just another slop generator — albeit a more fancy one.
When we refer to gen AI video as “slop,” we are usually commenting on aesthetics and presentation. But the means through which AI footage is created is a critical part of the equation. In contrast to traditionally produced movies, shows, and online videos — which can be sloppily crafted — things made with AI are “slop” because they are the products of workflows devoid of any direct authorial or artistic intent. Unlike a team of human filmmakers, a gen AI video model can’t always follow a story’s beats or a character’s motivations, but it can parse simple inputs and generate outputs that seem informed by a narrative (if you squint) because the program has been trained on vast amounts of visual data that is.
At its core, Seedance isn’t all that different from its peers
Being able to mimic the real (read: made by humans) thing is the entire point of projects like Seedance 2.0, but the models cannot do that unless they have first been given an ample amount of source material to programmatically iterate on. And by allowing such blatant IP infringement, ByteDance has told us that — aside from its zippier action shots and stronger sound design — at its core, Seedance isn’t all that different from its peers. It has been easy to recognize Seedance 2.0 as a slop generator when you look at the most viral clips that have been created with the program, which tend to feature A-list celebrities and obviously copyrighted fictional characters. But the trick of it all is much more difficult to suss out when you watch Chinese director Jia Zhangke’s Jia Zhangke’s Dance, a Seedance 2.0-generated short film featuring Zhangke having a debate about the nature of creativity with an AI version of himself.
Jia Zhangke’s Dance goes meta as its two characters discuss whether films created with AI should be thought of as gussied-up copies of human-made works or a new kind of art form. After one of the Jias reveals himself to be an AI copy of the other, the short follows them both on a Matrix-like journey through different settings that are meant to showcase AI’s ability to conjure up whatever images a prompter can think of. Jia Zhangke’s Dance unfolds with a smoothness and narrative cohesion you would be hard-pressed to find scrolling through OpenAI’s Sora app. But when you look closely at what’s going on in the background of the short’s busier scenes that involve background characters, it’s not hard to see Seedance 2.0 making some of the same continuity mistakes that plague all video generators.
Jia Zhangke’s Dance is a shining example of how filmmakers can make passable things with gen AI provided that they’re skilled enough to know how to work around the technology’s limitations. Though the film’s shots are very short, like most AI generated videos, they’re edited together in a way that creates the illusion of them being part of longer takes. And while characters in the distance will occasionally clip in and out of sight, you can see that Seedance 2.0 tries to obfuscate those mistakes by covering them up with objects moving in the foreground.
Filmmakers can make passable things with gen AI if they know how to work around the technology’s limitations
If anything, Jia Zhangke’s Dance shows us the degree to which many AI enthusiasts haven’t been trying especially hard to make their creations look like the kind of art that would put butts in theaters or get people to sign up for a streaming service. ByteDance’s engineers deserve at least some credit for building a model that can re-create real people’s faces with such accuracy. But it very much seems as if that strength might be tied to the model’s ill-gotten training data, which has put ByteDance into so much hot water that the company has paused its plans to release Seedance 2.0’s API to the public.
In addition to looking better than it currently does, the only way AI-generated video might be able to shed association with slop will be for the companies behind it to prove that their models can create things without needing to steal other people’s work. Studios like Asteria and companies including Adobe are trying to tackle that second issue with “IP-safe” models built with data that has been properly licensed. But until we start seeing quality work coming out of this new wave of AI programs, it’s going to be slop all the way down.
[Notigroup Newsroom in collaboration with other media outlets, with information from the following sources]
