OpenAI Launches Sora 2, AI-Generated Video Upgraded, "AI Version of Jitterbit" Opens New Creative Experience

October 1, early morning.OpenAIFlagship video and audio generation model releasedSora 2and launched the "AI version of Jitterbug" -SoraApps.

OpenAISora 2 is "heading straight for a GPT-3.5 moment in video". Compared to the previous Sora, Sora 2 is a more accurate and realistic simulation of the physical world, and is easier to control, with synchronized dialogue and sound effects.

Judging from the official video demo released, it can do some things that were difficult for previous video-generated models: Olympic gymnastics moves, accurately simulated buoyancy and backflips on paddleboards, and a three-and-a-half-week jump by a figure skater with a cat on her head.

▲Cue word: figure skater performs a triple half jump with a cat on her head

The Sora app is positioned as a social app that allows users to upload their own videos and participate in their friends' video creations through a "guest star" feature.OpenAI's goal is to try to create a new and unique way of interpersonal communication.

▲OpenAI Launches Social App Sora

The Sora iOS app is currently available for download now, launching on an invite-only basis in the US and Canada. Users who obtain an invitation code can also experience Sora 2 for free on the Sora website, and ChatGPT Pro users can experience the higher quality Sora 2 Pro model.

The release of Sora 2 exploded on the extranet, with most people asking for an invite code, and a small number of people already having one. Social platform X immediately appeared a lot of Sora 2 generation video, and even OpenAI CEO Sam Altman was "bad".

At the same time there are many who are concerned, with one user on X stating, "In a few months we won't be able to tell what's real and what isn't, and that will be a scary time."

▲Netizen Comments on Social Media Platform X

I. Sora 2 is here: a "GPT-3.5 moment in video"

According to OpenAI, the February 2024 release of Sora is in many ways the "GPT-1 moment" for video - video generation is starting to show results for the first time, and needs such as object persistence are being realized by scaling up pre-training computational power. realized through the expansion of pre-training computational power.

OpenAI calls Sora 2 "heading straight for a GPT-3.5 moment in video." Previous video models often deformed objects and distorted reality in order to successfully execute text prompts. For example, if a basketball player misses a shot, the ball may automatically travel to the basket. However, in Sora 2, if the basketball player misses a shot, the ball will bounce off the rim.

For example, the following video of a Sora 2-generated backflip shows the performer even stumbling a bit after landing and has a somewhat embarrassed look on his face from a small mistake, much like a real-life scenario.

▲Cue word: a man doing a backflip

Interestingly, the "mistakes" made by the model often seem to be internal to the implicit modeling of Sora 2.intelligent bodymistakes made; although it is still not perfect, it does a better job of following the laws of physics than previous systems.

OpenAI argues that this is an extremely important capability for any useful world simulator - you have to be able to simulate failure, not just success.

The model also takes a huge leap forward in controllability, being able to execute complex commands across multiple shots while accurately preserving the state of the world. It excels at handling realistic, cinematic and anime styles.

▲Cue word: Vikings at War - North Sea Launch (10.0 sec, cool winter daylight/early medieval) ......

As a general-purpose video and audio generation system, it is capable of creating complex background soundscapes, speech and sound effects with a high degree of realism.

▲ Cue word: two mountaineering explorers in brightly colored technical armor, faces frosted over, squinting, shouting eagerly through the snow, one at a time

Users can also inject real-world elements directly into Sora 2. For example, by looking at a video of one of our teammates, the model can insert him into any Sora-generated environment and accurately portray his appearance and voice. This feature is very versatile and applies to any human, animal or object.

▲Prompt word: Bigfoot was really nice to him, a little too nice, a little quirky. Bigfoot wants to play with him, but he wants to play too much.

OpenAI says the model is far from perfect and has many errors, but it confirms that further expansion of neural networks on video data will bring us closer to simulating reality.

Second, the AI version of Jitterbug launched, real people "cameo" video, the new social artifact?

Today, OpenAI also launched a new iOS social app called "Sora," powered by Sora 2.

In the app, users can create and remix each other's creative styles, discover new videos in customizable Sora dynamics, and introduce themselves or their friends to videos through the "Cameos" feature. With Cameos, users can bring themselves directly into any Sora scene with amazing fidelity by simply making a short audio or video recording in the app.

It looks like an AI version of Jitterbug or TikTok, and OpenAI believes that the social apps built around this "guest posting" feature are what make the Sora 2 experience so appealing.

A few months ago, OpenAI's team at Sora started experimenting with the "upload your own video" feature, and they're all having fun with it - OpenAI says it feels like a natural evolution in communication - from text messages to emoticons to voice memos and now to video. OpenAI says it feels like a natural evolution in the way we communicate - from text messages to emoticons to voice memos and now to video.

Last week, OpenAI released the app internally to all employees. Already, there has been feedback from coworkers that they have made new friends at the company through this feature.

Third, the invitation system is launched, Sora 2 is available for free, and the Pro user experience is more advanced

OpenAI has launched the Sora app on an invite-only basis to ensure that users can use it with their friends.

Upon receipt of an invitation, users will also be able to access Sora 2 through sora.com .Sora 2 will initially be available free of charge, but these features are still limited by computing power.ChatGPT Pro users will also be able to use the experimental, higher-quality Sora 2 Pro model on sora.com.

OpenAI also plans to release Sora 2 in the API. sora 1 Turbo will continue to be available and all user-created content will continue to exist in sora.com.

To prevent problems such as addiction, OpenAI will take a number of steps.

For one, it will provide users with the tools and autonomy of choice to take control of the content in their information stream. Utilizing OpenAI's existing large-scale language model, it has developed a new class of recommendation algorithms that can be guided by natural language; it also has built-in mechanisms to periodically survey users about their health and proactively provide them with options for adjusting their information flow.

By default, OpenAI displays content to users that is primarily targeted to people who follow or interact with it, and prioritizes videos that the model believes users are most likely to use as inspiration for their creations; it does not optimize for the amount of time a user spends in a dynamic stream of information, explicitly designing the app to maximize the amount of creations made, not the amount consumed.

On the teen protection front, OpenAI will launch Sora Parental Controls via ChatGPT so that parents can override infinite scroll limits, turn off algorithmic personalization, and manage private message settings.

In terms of the cameo feature, users have end-to-end control of the portrait with Sora. Only the user decides who can use their cameo, and can revoke access or remove any video containing it at any time. Users can view videos containing your cameo at any time, including drafts created by others.

OpenAI addresses many security issues in this app, such as informed consent for portrait use, provenance confirmation, prevention of harmful content generation, and more.

Many of the problems with other applications stem from their profitability models. openAI's only current plan is to eventually allow users the option to pay a certain amount to generate additional videos if the demand is too high relative to the available computing power.

Conclusion: Sora 2 Holds Back Big Moves, or Pushes Video Generation Industry Reshuffle

It's been over a year and a half since OpenAI released Sora in February 2024, and Sora 2 has finally arrived. Judging from the results, this model has made relatively great progress in terms of simulation realism, controllability and sound effects, and is expected to drive an accelerated reshuffling of the video generation industry landscape.

Video models are evolving at a rapid pace, and the Universal World Simulator not only offers new ways to generate content, but also promises to reshape interpersonal communication. openAI is getting closer to this goal with the new Sora social app, and marking a greater maturity of the video generation model in terms of on-the-ground applications.

Article source: Wisdom