You Can Now Try Out Sora, OpenAI’s AI Video Generator
Earlier this year, OpenAI, the creator of ChatGPT, announced Sora, an artificial intelligence video generator. Some of the demos the company showed off were hyper-realistic, from puppies playing in the snow to the view from a subway car driving through the city. In short, it was both impressive and terrifying, as I explained in my initial thoughts here .
Sora news has been largely silent since then: the company hasn’t provided a platform for public testing, so we don’t have much personal experience to work with. This week the situation will change. Not only is OpenAI making Sora more accessible, but it’s also opening up its AI video generator to anyone with a ChatGPT Plus account. Things are about to get wild.
What’s new in Sora
OpenAI made the announcement on Monday following an earlier report from Marques Brownlee . If you followed Sora’s initial announcement closely, none of the examples given here will shock you: In short, Sora may be asked to create photorealistic short videos in a variety of situations: OpenAI and Marques Brownlee have demo shots of drones. rocks, animals in nature and people performing tasks “on camera”. But what’s really new today is a number of features that OpenAI has added to the Sora program as a whole.
There is “Storyboard”, a kind of video editor that allows you to combine different video prompts to create longer videos on the same topic. For example, you might have one prompt asking for a crane to stand in the water, and another prompt asking for the crane’s head to be placed in the water. Sora will then combine the two clues into one continuous video. Recut also acts as a video editor, only here you can isolate a specific part of your video and ask Sora to expand it. “Remix” opens a new prompt field that allows you to request changes to an existing video. (You can also choose the “strength” of the remix, which affects how much of the video will actually be changed based on your request.) Finally, Blend allows you to turn the theme of one video into another. Sora’s example: the butterfly from one video turned into an orchid from the second video.
Of all these new features, Storyboard seems to be the most interesting. This seems to be a clever workaround to allow the AI to generate a complex scene with multiple actions, since trying to cram it all into one clue would likely fail. Remixing can also theoretically be useful for fine-tuning elements of a video without having to throw away the original generation. But at the end of the day, our collective interest in the model stems from its primary function: you give Sora an idea for a video, and she generates it for you. Or you upload a photo from your library and Sora brings inanimate objects to life into a moving scene.
Sora in action
When you submit a request, your video is added to a “queue” for processing. The time it takes to create a video depends on your settings, including resolution, duration, and the number of variations you create: I’m on the standard ChatGPT Plus plan, so I’m limited to a maximum resolution of 720p and a maximum duration. five seconds. ChatGPT Pro users can increase this resolution to 1080p, create videos up to 20 seconds long, and create up to four variations of their video.
Unfortunately for me, it seems like everyone, including their mother, is trying to take advantage of Sora now. My first and only operational attempt (“filming a taxi driving through the city center”) was hanging in limbo the entire time I was writing this article. In fact, OpenAI has suspended account creation for now because there are too many people trying to access the video generator.
My video finally generated and it was rough. The video quality of the taxis and the city was quite good (again, very photorealistic), but the taxi movements were all over the place. He first drove in reverse, then transformed into a car going in the other direction, and then sped away as another taxi pulled into the foreground. (My old taxi also disappeared into thin air, and the new taxi did not have a trunk, but two front ends.)
Since Sora takes such a long time to generate videos, it’s useful now to look at someone like Marques Browni, who has spent some time testing the tool. In his review of Sora, he finds that the tool still has difficulty avoiding the common pitfalls of AI-generated videos : videos may look photorealistic, but lose realism in motion. Sora often confuses which leg should be in front and which should be behind when walking, or generally “forgets” about objects. When Brownlee asked for a video of a tech reviewer covering a smartphone, the reviewer is holding two smartphones and one just disappears for no reason. Some aspects of the video may play in slow motion while others run at normal speed, which seems odd at first glance. These glitches are common in most Sora products I’ve seen: if you look for them, you’ll see them, and they call attention to the artificiality of the video.
This is also true for “low quality” videos such as CCTV footage or CCTV footage. Cars crash into each other and disappear, or people move in unrealistic ways. But I will say that the low quality of these videos makes it easy to fake: if Sora can figure out the physics, people will spend all day inventing CCTV footage that doesn’t exist.
In Brownlee’s experience, what Sora currently does best is not realistic at all: for example, the motion graphics, like some clips of animation and animated characters, usually look good. For example, the animation of the Empire State Building looks like something out of the intro to a Netflix series. And when Brownlee uploaded an image of animated leaf characters created by DALL-E, Sora animated the image in a fairly believable manner. It’s a little easier to ignore the flaws when the video doesn’t pretend to be real at all.
Sora also does a good job with drone composing and tracking shots, with a drone shot of Mount Fuji or the Golden Gate Bridge looking smooth and photorealistic. If you look closely, you might notice glitches and imperfections, such as waves not behaving quite as they should, but you could probably insert that footage into shows and movies without many (or most) people noticing. noticed.
Where should we go next?
Sora scared me back in February when it was announced. It’s been ten months since then and I’m still scared, but not because the videos have gotten so much better. In fact, from what I see today, the quality seems to be about the same, albeit with some new AI features that you can use to customize these videos. The realism is still there when it is, as are the flaws, of which there are many.
What scares me is the availability: once OpenAI meets demand, Sora will be available to anyone with a ChatGPT Plus subscription. For $20, you get access to a tool that can create up to 50 five-second videos per month. Five seconds certainly isn’t that long, so without some savvy these videos likely won’t do the most damage.
This is where ChatGPT Pro comes in. This plan is much more expensive ($200 per month), but for that $200 you can create up to 500 videos, each of which can be up to 1080p resolution and up to 20 seconds long. OpenAI says you can also download these videos without a watermark, which will make detection much more difficult.
Of course, most of us wouldn’t sign up for Pro for this, but $200 isn’t much of a deterrent to attackers looking to spread misinformation. Imagine the next major polarization crisis, caused by a flood of videos that somehow “prove” what happened, when in fact those videos are not real at all. OpenAI does have some security features built into it, such as blocking videos from including copyrighted material or famous figures, but we’ll see how well these checkpoints work in practice.
How to try Sora
Account creation for Sora is currently not available, but that may change soon. If you’d like to try Sora for yourself, go to sora.com . From here, click “Login” and then authenticate yourself using your ChatGPT account. Remember that to use Sora you will need either a ChatGPT Plus account ($20 per month) or a ChatGPT Pro account ($200 per month).