NewsTech

OpenAI Announces Big Changes to Its Sora AI Video Generator

OpenAI’s new Sora video generator is less than a month old, and remains available only by invitation.

Still, OpenAI is already tweaking it. Bill Peebles of OpenAI announced on X yesterday that the company is making some big changes to Sora.

Here’s what’s coming, and why it matters.

Deepfaking Your Dog

One of Sora’s most popular features is called “Cameos.” Basically, you let Sora briefly film you and record your voice, and then the app can insert you into your own AI-generated videos.

In my testing, I’ve found Cameos to be hit or miss. Some peoples’ Cameo looks just like them, while others are way off.

Personally, my Cameo is far larger than me. That’s probably because Sora looks only at the user’s head and face when creating their Cameo, and then infers the shape and appearance of their body.

I have a very large head. Therefore, Sora assumes my body is large, too. In fact, my real-life head just doesn’t quite fit my body. AI isn’t perfect…

Now, Sora is improving Cameos by allowing you to scan yourself, but also to scan favorite objects, and even pets.

If you want to deepfake your dog (something I’ve done continuously with my dog Lance since AI video generators rolled out) or make a video where your kid’s favorite stuffie comes to life and ransacks a village, you’ll be able to.

Making Movies

Currently, Sora’s clips are short — about 11 seconds. That’s pretty limiting in terms of what you can do creatively.

There’s also no way to edit or tweak them. Sometimes a clip starts out fine, but then it goes off the rails.

Now, Peebles says you’ll be able to edit clips, combining them together to create longer movies, or presumably editing out parts that didn’t turn out right.

Peeble also says “very powerful new features” for editing will roll out soon.

More TikTok-y

Beyond Cameos and editing, Peebles says that OpenAI is making changes to improve the Sora feed, to make it more social, and to make its UI snappier.

Basically, it sounds like OpenAI wants to make Sora more like TikTok, which has strong social features (friending, seeing content from people you follow, etc.), and has a very snappy interface. You’ll apparently be able to see Sora videos from your contacts, but also to join groups related to workplaces, schools or other institutions.

Below the surface, OpenAI is likely improving the recommendation algorithm that chooses which videos to show to specific users. This type of algorithm is what makes TikTok so powerful, so OpenAI probably wants to duplicate their success here, too.

Bigger picture, it seems like OpenAI rolled out Sora without really expecting how well it would do. Now that the app’ is dominating Apple’s App store, they’re doing the core work to make it grabbier, faster and more socially engaging.

Those Pesky Admins

Users have noticed that Sora has a tendency to moderate videos fairly aggressively. Describing this as “super annoying”, Peebles says that OpenAI is working to adjust it and allow more videos through.

That’s a bit surprising, as Sora has already been nailed for duplicating trademarked or copyrighted characters, and creating offensive deepfakes of Martin Luther King Jr.

Hopefully the company can walk a better line between yanking videos that are harmless, and allowing through this kind of damaging (and lawsuit-provoking) slop.

In addition to those changes, OpenAI says it will soon release an Android version of Sora, and asked users to continue giving the company feedback.

Thomas Smith

Thomas Smith is a food and travel photographer and writer based in the San Francisco Bay Area. His photographic work routinely appears in publications including Food and Wine, Conde Nast Traveler, and the New York Times and his writing appears in IEEE Spectrum, SFGate, the Bold Italic and more. Smith holds a degree in Cognitive Science (Neuroscience) and Anthropology from the Johns Hopkins University.

Leave a Reply

Back to top button

Discover more from Bay Area Telegraph

Subscribe now to keep reading and get access to the full archive.

Continue reading