New slop just dropped, from OpenAI
How many wings does an airplane have? OpenAI's video generator doesn't know.
OpenAI, the company behind the ChatGPT chatbot, DALL-E image generator, and GPT large language models (LLM), just released another showcase of its AI video generator Sora. It's shockingly, but perhaps not surprisingly, awful.
OpenAI announced Sora back in February as a text-to-video diffusion model. It's supposed to be able to create videos from simple text prompts, much in the same way DALL-E, Microsoft Creator, Midjourney, and other similar tools can already create images from text prompts. The company said at the time, "Sora can generate videos up to a minute long while maintaining visual quality and adherence to the user’s prompt."
The initial demonstration clips of Sora were eerie and low-resolution, with bizarre movement and physics. The video generation model was almost certainly trained on mountains of stolen content and sucked up excessive amounts of power and water. The most impressive sequences were nature shots, but that was probably because they were close to the model's training material. It was an initial demonstration, though, so maybe OpenAI just needed more time to work on it.
Fast forward to October 2024. OpenAI is now valued at $157 billion: more than the national GDP of Saudi Arabia, Turkey, or Indonesia. ChatGPT has "more than a hundred million" weekly active users, and the underlying GPT language model is used by many other services, such as Microsoft Copilot and some answers in Apple's revamped Siri. Sora is still not publicly available, but OpenAI is giving a few people access to the in-development version.
OpenAI just published a new 'Dreams' short film on its YouTube channel. It's a reading of a poem written by IN-Q on top of a sequence of Sora-generated video clips, produced in collaboration with filmmaker Wayne Price.
The video is awful. It's one of the best representations of "AI slop" I have seen, except it's not coming from a weird Facebook page. It's live right now on the YouTube channel of OpenAI, one of the biggest companies in the AI space.
A fever dream
The 'Dream' video is nearly two minutes long, showing back-to-back clips of a bird flying around airplanes, while a recording of a poem plays in the background about the bird wishing to be an airplane. OpenAI does have a text-to-speech technology—one of the voices sounded suspiciously like Scarlet Johansson, after the actor denied requests to license her voice—but the poem recording seems to be a human voice. It's identical to the track in the artist's album release from earlier this year.
Nearly every shot in the film has a significant flaw in the background, objects, and physics. The bird that appears in most scenes is also changing colors, even though I think it's supposed to be the same bird throughout the runtime, as the poem is told from its perspective. Wayne Price explains in the video description that the film is supposedly representative of "hyper reality with a small dose of the fantastical, and that "it took many attempts to get the gold."
The video starts with some shots of the bird flying near an airport, then it cuts to the bird high up in the air next to a plane. The plane's wing is parallel to the main body for some reason, instead of connected to the side.
A few seconds later, the bird is flying next to another plane. It passes one wing, then passes... another wing on the same side. Having some trouble there.
The very next cut shows the bird next to a (seemingly) landed airplane, with the side door and staircase open. The entire plane behind the front section is missing, replaced with a wing and jet engine. Maybe this was just a plane built by Boeing?
There are a few cuts of the bird flying around the cockpit, then a fisheye lens-style view of a plane making an air trail. This plane has wings with tips that are tilted upwards for some reason.
A few scenes later, there's a close-up view of the bird flying with a plane barely visible in the background. This plane has a third wing.
There are a few more shots of the bird flying around that are mostly decent, then one where it's sitting on a pile of toy planes in a marketplace. The planes are all morphed into each other, and the bird now has a completely red face.
A while later, the bird is once again flying near an airport runway. This time, the plane appears to have a normal number of wings, but one of the wings has an extra jet engine.
The last shot shows the bird turning into a plane, Animorphs-style. Price described this as the "key moment" of the short film to "coincide with the final line of IN-Q’s brilliant poem." It's a neat effect.
It's not entirely clear how many of the errors here were intended as "a small dose of the fantastical." That could have just as easily been an artistic choice as it could have been Price giving up after typing "orange bird" in the video generator prompt for the 50th time and still getting a red bird.
What is clear is that this video serves no purpose. It fails at its mission at being a technical demonstration of Sora video generation, because even the elements that are (probably) supposed to look realistic don't look real or move realistically. It's not really interesting as a creative work, because it's just jump cut after jump cut of a bird flying around misshapen airplanes.
If the goal was to showcase the poem in a visual medium, then simple stock footage of birds and airplanes would have been less distracting. Instead, it degrades the entire work to the point where some at least some viewers thought the entire thing was AI-generated.
Art is subjective, but this isn't really art, it's just AI slop. It’s incredible to me that this is what OpenAI thinks we should be excited about.
