Last February, OpenAI teased a powerful AI generator dubbed Sora that could create complete videos from simple text prompts. Today, Sora leaves beta and is launching to the public.
Available as part of ChatGPT subscriptions starting at $20/month, users will be able to create high definition videos up to 20 seconds long from nothing more than a text prompt. However, the capabilities and UX of Sora are both far beyond the spartan presentation of ChatGPT.
Most notably, a small pool of pilot testers will be able to upload images of people to place real humans into imaginary scenes (a feature that OpenAI says it may or may not release to the full public).
“We’re running the experiment for a certain amount of time before making the call,” says Rohan Sahai, product lead on Sora.
Subscribe to the Design newsletter.The latest innovations in design brought to you every weekday
[Image: OpenAI]Inside the design of SoraWhat surprised me most about my remote demo of Sora was simply how articulated the early interface seems to be—and how much richer its entire approach to UX is compared to the launch of ChatGPT a couple years ago. The company is striking a balance between a powerful, customizable AI workflow and something simple enough for the average person to understand.
[Image: OpenAI]“We’ve needed to add new interfaces,” says Souki Mansoor, artist program lead on Sora.
The prompt box is just one of a few options to generate videos. You can also import your own inspiration media, and even train your own stylized aesthetic. The interface greets you in a file management system that could be straight out of Google Drive. You have a directory of projects on the left, and your workspace on the right.
On the very top of the list is a feed, which includes recent works by the Sora community, presented as a stoic mood board reminiscent of visual design depositories. “We were thinking of it as less a social media feed than inspiration and education, to help people know how to use Sora in more effective ways,” says Mansoor. Tap into any of those videos, and you can not only watch it, but see the prompts and methods behind it, and you can even remix that video (sort of like AI TikTok) if you like.
But when you actually get into creating your own videos, the larger tool set looks something like Adobe Premiere or iMovie. The moment you start generating a video, you can hop over to a timeline view, where Sora has outlined what it is generating for you in a storyboard of text cards.
[Image: OpenAI]How Sora worksI’ve tasked the team with generating an axolotl on the beach. Even in this early storyboard, I can see how OpenAI has enriched my text prompt to ensure I have a video worth watching, describing the animal as pink with “distinctive feathery gills” with “the gentle waves lapping as they shimmer.” In the next storyboard panel, I learn what the axolotl is doing, since I didn’t give it a task. It “moves slightly,” with fluttering gills as it looks toward the ocean.
[Image: OpenAI]I could hop into these storyboards to rewrite them. Or I could hop into a bottom navigation bar to change the resolution, or add a stylistic filter of my own trained creation, like film noir. When I ask how much power each of these videos take to generate, OpenAI offers the equivalent of a “no comment.”
[Image: OpenAI]The axolotl I got had too many gills, but it was otherwise a convincing natural video. Then the team demonstrated its other video timeline tools, like blending two scenes together, lengthening the video to just be longer, or sampling another clip to inspire the aesthetic of your own.
These capabilities were developed over the last year in a closed beta OpenAI has been running with creatives across the world. One of those groups was the Canadian music group Shy Kids, which produced a genuinely affective video for their latest single, “My Love,” in which two macaque monkeys bask in their companionship in an unspoiled rainforest . . . before a sad reveal. It’s just the sort of imagery you could spend a lifetime chasing as a wildlife photographer and never witness. But Shy Kids produced the video virtually in two weeks with no VFX outside Sora other than light editing and color correction.
VIDEO
“It’s very good at ensuring there is never a blank page. It allows you to ideate and iterate quickly, which allows more people into the creative process,” writes group member Walter Woodman via email. “It still needs work on specificity. That being said you don’t get upset at a toddler for not being able to understand physics . . . yet!
“Our plan is to continue to explore its boundaries . . . and not,” he continues. “I am currently in Marrakech, the guy who lugs around a 10-pound analog camera and film because it is so important to make images the old way. That being said—also extremely important to explore and shape the new.”
[Images: Minne Atairu/OpenAI]Stress testing the system, and deciding the future of deepfakesWhile OpenAI has been working with artists for the last year, the company has really been living two lives. “The artists are not malicious users,” notes Sahai. So to stress test the system, OpenAI turns to its “red teamers,” who attempt to root out every terrible use case before the public can.
OpenAI describes the protections behind Sora as its “safety stack.” Much of it simply builds off all of the work OpenAI has already done with ChatGPT and its image generator Dall-E. ChatGPT language analysis scrubs through every request. You can’t, for instance, generate nudity or violent imagery on Dall-E today.
Their goal is to “balance creativity and misuse,” says Sahai. “We’re starting with a fairly cautious approach. We want to be extremely conservative for anything in the realm of child sexual abuse, or revenge porn deepfakes, [so these] use cases to become essentially impossible on Sora. I expect a lot of creatives will say it’s a little too trigger happy on refusals, and we’ll be figuring out that balance over time.”
I was the first journalist to interview the anonymous developer behind the original deepfake method back in 2018, and while the work was mostly laborious and crude, it was clear that we’d entered a radical new chapter in humanity, when anyone could be rendered into any context more or less instantly. Six years later, Sora has brought us into that future.
The easiest option to mitigate the worst abuse with Sora would be to ensure that no recognizable human could be generated inside of it. As the team explains, however, many artists expressed interest in inserting particular people into scenes from sample imagery—and to be fair, anyone in creative production could really use such a tool.
“I come from a background in film…so [I understand wanting] Sora to support videos that include real people,” says Mansoor. “We want to avoid misuse.”
So what are the limiters to misuse? OpenAI is launching human generation through that aforementioned limited beta. It’s promising that human moderators will study outputs closely, and using visages of celebrities and public figures will be blocked by default. All produced videos are invisibly watermarked for trackable provenance. They’ll also ban anything that verges into a “pattern of misuse” or IP infringement.
“We’re treating this as an experiment,” says Sahai. “It’s just a certain amount of people we can closely monitor.”
Personally, I’m less worried about the possibility of AI revenge porn, which feels mitigable with enough protections, than I am my own lack of consent to be placed into someone else’s imagined meme. I suggest that a more responsible approach, if OpenAI were focused on doing human generation ethically, would be to introduce stricter opt-in and opt-out options.
With an opt-in feature, artists could only generate videos of themselves, for instance. Or perhaps anyone could simply go onto OpenAI and opt-out of having their image used. The hypothetical option of opting out would be logistically complicated as it would require OpenAI to have to study every face coming into their system.
“This is something we’re going to consider,” says Sahai on the suggestion of opt-out tools. “There’s a lot of biometric complexity if we store face data and tools on top of that, and does this become such a barrier where…it’s easier just to shut off the feature.”
OpenAI’s Sora is available now. Reality is gone forever.
The extended deadline for Fast Company’s World Changing Ideas Awards is this Friday, December 13, at 11:59 p.m. PT. Apply today.
{Categories} _Category: Implications{/Categories}
{URL}https://www.fastcompany.com/91243136/openai-sora-public-launch-deepfakes{/URL}
{Author}Mark Wilson{/Author}
{Image}https://images.fastcompany.com/image/upload/w_1280,q_auto,f_auto,fl_lossy/q_auto,c_fit/wp-cms-2/2024/12/p-3-open-ai-sora.gif{/Image}
{Keywords}Design{/Keywords}
{Source}Implications{/Source}
{Thumb}{/Thumb}