Sora 2 is here

今天我们发布了 Sora 2 —— 我们的旗舰视频与音频生成模型。

最初的 Sora 模型（2024 年 2 月） （https://openai.com/index/video-generation-models-as-world-simulators/）在许多方面相当于视频领域的 GPT‑1 时刻——这是视频生成首次开始显得“可行”的时候，通过扩大预训练计算量，像物体永久性这样的简单行为开始出现。从那以后，Sora 团队就专注于训练具有更高级世界模拟能力的模型。我们认为，此类系统对于训练能够深入理解物理世界的 AI 模型至关重要。实现这一目标的一个重要里程碑，是掌握在大规模视频数据上进行预训练和后训练，而与语言相比，这方面还处于初期阶段。

有了 Sora 2，我们直接跳到我们认为可能是视频领域的 GPT‑3.5 时刻。Sora 2 能做到许多对先前视频生成模型来说极其困难——在某些情况下甚至是不可能的事情：奥林匹克体操动作、在桨板上完成的后空翻并能准确模拟浮力和刚性的动力学、以及当一只猫拼命抓着不放时的三周轴动作。

示例提示：a guy does a backflip （示例提示翻译：一个男人做后空翻）

先前的视频模型往往过于乐观——它们会为了满足文本提示而扭曲物体或变形现实。例如，如果一名篮球运动员投篮未中，球可能会自发地传送到篮筐里。在 Sora 2 中，如果篮球投失了，它会从篮板上反弹。有趣的是，模型所犯的“错误”往往看起来像是 Sora 2 内部隐式建模的智能体的错误；尽管仍不完美，但与先前系统相比，它更能遵守物理定律。这对任何有用的世界模拟器来说是一个极为重要的能力——你必须能够模拟失败，而不仅仅是成功。

该模型在可控性方面也有巨大进步，能够遵循跨多镜头的复杂指令并准确地保持世界状态。它在现实主义、电影级和动漫风格上表现尤为出色。

示例提示：intense anime battle between a boy with a sword made of blue fire and an evil demon demon （示例提示翻译：一场激烈的动漫战斗：一个男孩持有由蓝色火焰构成的剑，与一个邪恶恶魔对战）

作为一个通用的视频-音频生成系统，它能够创建复杂的环境音景、语音和逼真的音效。

示例提示：Two mountain explorers in bright technical shells, ice crusted faces, eyes narrowed with urgency shout in the snow, one at a time （示例提示翻译：两名穿着明亮高性能外壳的登山者，脸上结着冰，双眼紧盯，急切地在雪地里轮流呼喊）

你还可以将现实世界的元素直接注入到 Sora 2 中。例如，通过观察我们一位同事的视频，模型可以将他们插入到任何 Sora 生成的场景中，并准确还原其外貌和声音。此能力非常通用，适用于任何人类、动物或物体。

示例提示：Bigfoot is really kind to him, a little too kind, like oddly kind. Bigfoot wants to hang out but he he wants to hang too much （示例提示翻译：大脚怪对他非常友好，有点过分友好，奇怪地友好。大脚怪想要一起玩，但他想得太多了）

该模型还远未完美，会犯许多错误，但它验证了在视频数据上进一步扩大神经网络规模将使我们更接近模拟现实。

SORA 2 的部署

在通向通用世界模拟和能够在物理世界中运作的 AI 系统的道路上，我们认为人们可以在我们构建的这些模型上获得很多乐趣。

几个月前，Sora 团队内部就开始玩这个“上传自己”功能，我们都玩得很开心。它某种程度上感觉像是沟通方式的自然演进——从短信到表情符号、到语音信息，再到现在。

因此今天，我们推出了一款新的社交 iOS 应用，名为“Sora”，由 Sora 2 提供技术支持。在应用内，你可以创建并改编他人的生成作品、在可定制的 Sora 动态中发现新视频，并通过“客串”（cameos）将你自己或朋友带入生成场景。使用客串功能，你可以通过在应用中进行一次短时间的视频和音频录制来验证身份并捕捉肖像，从而以极高保真度把自己直接放入任何 Sora 场景。

上周，我们在 OpenAI 内部向所有员工推出了该应用。我们已经收到同事们的反馈，说他们因为这个功能在公司结识了新朋友。我们认为，围绕这个“客串”功能构建的社交应用是体验 Sora 2 魔力的最佳方式。

负责任地推出

关于刷屏（doomscrolling）、成瘾、孤立以及强化学习优化的推荐流（RL‑sloptimized feeds）等担忧，我们非常重视——以下是我们的应对措施。

我们为用户提供工具和选择权，让他们掌控动态中看到的内容。借助 OpenAI 现有的大型语言模型，我们开发了一类可以通过自然语言指示的推荐算法。我们还内置机制，定期询问用户的身心状况，并主动给予他们调整动态流的选项。

默认情况下，我们会向你展示严重偏向你所关注或互动对象的内容，并优先展示模型认为你最有可能将其用作自己创作灵感的视频。我们并不以用户在动态中停留时间为优化目标，并且明确将应用设计为最大化创作而非消费。你可以在我们的动态哲学（Feed Philosophy）中找到更多细节（https://openai.com/index/sora-feed-philosophy/）。

这个应用是为与朋友一起使用而设计的。测试者给出的压倒性反馈是，正是客串功能让这一体验不同且有趣——你必须亲自试试才能真正体会，但它是一种新颖且独特的沟通方式。我们将以邀请制方式推出，以确保你能与朋友一起进入。在所有主流平台都在远离传统社交图谱的当下，我们认为客串功能将强化社区纽带。

保护青少年的身心健康对我们很重要。我们对青少年在动态中每天可看到的生成内容默认设置了上限，同时对该群体的客串权限也实施了更严格的限制。除了我们的自动化安全体系，我们还在扩充人工审核团队，以便在出现欺凌案件时快速处理。我们将通过 ChatGPT 推出 Sora 家长控制（https://openai.com/index/introducing-parental-controls/），家长可以通过它覆盖无限滚动限制、关闭算法个性化，以及管理私信设置。

对于客串，你可以对自己的肖像在 Sora 中进行端到端控制。只有你决定谁可以使用你的客串，你可以随时撤销访问或删除包含你客串的任何视频。包含你客串的视频（包括他人创建的草稿）你可以随时查看。

我们在该应用上处理了许多安全话题——关于使用肖像的同意、来源可追溯、阻止生成有害内容等更多内容。详见我们的 Sora 2 安全文档（https://openai.com/index/launching-sora-responsibly/）。

许多其他应用的问题源于其货币化模式在激励上与用户身心健康相冲突。开诚布公地说，我们当前唯一的计划是：如果对生成的需求相对于可用算力过大，最终可能给用户提供付费选项，以便生成额外的视频。随着应用的发展，我们会公开沟通任何方法上的改变，并持续把用户身心健康作为我们的主要目标。

我们才刚刚开始这段旅程，但凭借 Sora 2 带来的强大创作与改编方式，我们认为这是共创体验全新时代的起点。我们乐观地认为，这将比现有平台为娱乐与创造提供更健康的环境。希望你玩得愉快 :)

SORA 2 的可用性与未来规划

Sora iOS 应用（在新窗口中打开）（https://apps.apple.com/app/id6744034028）现在可以下载。你可以在应用内注册以便在你的账户开放访问时收到推送通知。我们今天在美国和加拿大开始初步推出，计划迅速扩展到更多国家。收到邀请后，你也可以通过 sora.com（在新窗口中打开）（http://sora.com/）访问 Sora 2。Sora 2 初期将免费提供，并且一开始会有慷慨的使用额度，让人们可以自由探索其能力，但这些仍受算力限制。ChatGPT Pro 用户也将能在 sora.com（在新窗口中打开）（http://sora.com/）上使用我们实验性的更高质量 Sora 2 Pro 模型（并且很快也将在 Sora 应用中提供）。我们也计划在 API 中发布 Sora 2。Sora 1 Turbo 将继续可用，你所创建的一切内容将继续保存在你的 sora.com（在新窗口中打开）（http://sora.com/）库中。

视频模型进步飞快且迅速。通用世界模拟器和机器人代理将从根本上重塑社会并加速人类进步的步伐。Sora 2 代表了朝这个目标迈出的重要一步。根据 OpenAI 的使命，随着这些模型的发展，让全人类受益是至关重要的。我们认为 Sora 将为世界带来许多欢乐、创造力与联系。

—— Sora 团队 撰文

SORA 2 团队名单（节选）

研究（Research） Harold Li, Dmytro Okhonko, Avi Verma, Eric Zhang, Ricky Wang, Troy Luhman, Eric Luhman, Bram Wallace, Eric Mintun, Michael Chang, Gabriel Petersson, Jure Zbontar, Daniel Geng, Will DePue, Alex Zhao, Cheng Lu, Yufei Guo, Pritam Damania, Larry Kai, Farzad Khorasani, Kenji Hata, James Betker, Vladimir Chalyshev, Connor Holmes, Aditya Ramesh, Bill Peebles

产品（Product） Andrew Kondrich, Andrew Sima, Andrew Thieck, Andrey Malevich, Antonio Di Francesco, Bin Wen, Bing Liang, Boyang Niu, Cheng Su, Cristina Scheau, Daniel Latta-Lin, David Schnurr, Dhruba Borthakur, Duc Tran, Gilman Tolle, Greg Hochmuth, Joe Taylor, Joey Flynn, Joey Pereira, Julius Hochmuth, Key Shin, Liam Esparraguera, Liang Wu, Liang Xiong, Mengchao Zhong, Michelle Hwang, Mick Jermsurawong, Mike Starr, Omar Elfanek, Patrick Hum, Pavel Komlev, Rajeev Nayak, Raunak Daga, Rohan Sahai, Sergii Rudenko, Shuyi Chen, Tarek Younes, Thomas Bredillet, Thomas Bredillet, Thomas Dimson, Victoria Huang, Vladimir Chalyshev, Welton Wang, Wesam Manassra, Xiaolong Wang, Yizhe Yu, Yun Jiang, Zhigang Wang

贡献者（Contributors） Aarash Heydari, Chad Nelson, Daniel Fradin, David Duxin, Hessam Bagherinezhad, Jasmyn Samaroo, Jay Wang, Jess Manzano, Kendra Rimbach, Nikki Sommer, Sergei Vorobev, Shirong Wu, Soham Govande, Souki Mansoor, Tifa Chen, Tomer Kaftan, Tyce Walters, Varun Shetty

领导（Leadership） Bill Peebles — Sora Connor Holmes — Systems Rohan Sahai — Product Thomas Dimson — Product Natalie Summers — Chief of Staff Aditya Ramesh — Organization

特别感谢（Special Thanks） Adam Majmudar, Adele Li, Aravind Suresh, Arun Vijayvergiya, Ashkay Pall, Ben Leimberger, Brad Lightcap, Charlotte Cole, Chris Hallacy, Chris Koch, Christine McLeavey, Christopher Lehane, Dane Stuckey, Eric Wallace, Fidji Simo, Gabriel Goh, Gary Briggs, Geoff Salmon, Giancarlo Lionetti, Greg Brockman, Hannah Wong, Ian Sohl, Jakub Pachocki, Jamie Kiros, Jason Kwon, Jeffrey Han, Joanne Jang, Johannes Heidecke, Josh Achiam, Kate Rouch, Kevin Weil, Lauren Itow, Li Jing, Mark Chen, Mark Gewurz, Matt Knight, Matthew Isono, Max Burkhardt, Mayank Gupta, Mia Glaese, Nick Turley, Patrick Geonetta, Peter Welinder, Philip Bogdanov, Prafulla Dhariwal, Robert Xiong, Ryan O'Rourke, Sam Altman, Sarah Friar, Sarah Russell, Sarah Warkov, Specer Papay, Srinivas Narayanan, Sulman Choudhry, Szymon Sidor, Tejal Patwardhan, Vikki Lampton, Vlad Fomenko, Wojciech Zaremba, Young Cha, Yuchen Zhang

安全、诚信、产品政策、i2、用户运营（Safety, Integrity, Product Policy, i2, User Ops） Adam Wells, Aleah Houze, Annie Cheng, Artyi Xu, Carolina Paz, Claudia Fischer, Garrett Harkins, Gilman Tolle, Jackie Hehir, Jake Brill, Jesika Haria, Kate Birks, Kelly Stirman, Lauren Jonas, Mentong Zhang, Pedram Keyani, Pedro Aguilar, Ryan Rinaldi, Sam Toizer, Sarah Ryan, Savannah Heon, Shalli Jain, Shauna O'Brien, Tim Boll, Zoe Stoll

法务（Legal） Tyce Walters, Ali Buttars, Brian McKnight, Gideon Myles, Tom Rubin, Dani Westbrook, Charles Proctor

传播（Communications） Alex Baker-Whitcomb, Anna McKean, Ashley Tyra, Bailey Richardson, Gaby Raila, Julie Steele, Leah Anise, Niko Felix

市场、设计与创意（Marketing, Design, & Creative） Adam Brandon, Adrian Gunadi, Alexandr Khomyakov, Anne Oburgh, Antonia Richmond, Ben King, Cary Hudson, Chloe Bowers, Chris Hutchinson, Ciaran Rogers, Dalhae Lee, Dana Palmié, Daniel Stuhlpfarrer, Daniel Zhang, Elisha Greenwell Dargan, Ian Silber, Indgila Sama Ali, Jeffrey Sabin-Matsumoto, Josh Cleveland, Kaitlin Giannetti, Kenneth Kuh, Kim Baschet, Malisa Kuch, Melia Tandiono, Michaela McCrink, Minnia Feng, Nick Ciffone, Paymon Parsia, Phillip Kim, Phillip Kim, Raegan Allsbrook, Roy Chen, Shannon Jager, Thomas Degry, Xingle Huang, Yara Khakbaz, Zach Stubenvoll

全球事务（Global Affairs） Claudia Fischer, Debbie Mesloh

战略财务（Strategic Finance） Chengpeng Mou, Caroline Zhao

API Adam Wells, Alina Wu, Amelia Liu, Andi Liu, Ankit Gohel, Annie Cheng, Artyi Xu, Brian Ratajczak, Chad Nelson, Erika Kettleson, Filippo Raso, Gilman Tolle, Jackie Hehir, Jeff Harris, Jen Robinson, Joanne Shin, Jono Oko, Katia Gil Guzman, Kelly Stirman, Leher Pathak, Manoli Liodakis, Miqdad Jaffer, Olivia Morgan, Robin Koenig, Rohan Sahai, Ruth Costigan, Sarah Ryan, Savannah Heon, Shaokyi Amdo, Shaili Jain, Tabarak Khan, Tonia Osadebe, Tyce Walters, Wei Sun, Wendy Jiao, Woo Kim, Yi Ma

由 OpenAI 在加利福尼亚州旧金山构建 发布日期：2025 年 9 月 30 日

Today we’re releasing Sora 2, our flagship video and audio generation model.

The original Sora model⁠ from February 2024 was in many ways the GPT‑1 moment for video—the first time video generation started to seem like it was working, and simple behaviors like object permanence emerged from scaling up pre-training compute. Since then, the Sora team has been focused on training models with more advanced world simulation capabilities. We believe such systems will be critical for training AI models that deeply understand the physical world. A major milestone for this is mastering pre-training and post-training on large-scale video data, which are in their infancy compared to language.

With Sora 2, we are jumping straight to what we think may be the GPT‑3.5 moment for video. Sora 2 can do things that are exceptionally difficult—and in some instances outright impossible—for prior video generation models: Olympic gymnastics routines, backflips on a paddleboard that accurately model the dynamics of buoyancy and rigidity, and triple axels while a cat holds on for dear life.

Prompt: a guy does a backflip

Prior video models are overoptimistic—they will morph objects and deform reality to successfully execute upon a text prompt. For example, if a basketball player misses a shot, the ball may spontaneously teleport to the hoop. In Sora 2, if a basketball player misses a shot, it will rebound off the backboard. Interestingly, “mistakes” the model makes frequently appear to be mistakes of the internal agent that Sora 2 is implicitly modeling; though still imperfect, it is better about obeying the laws of physics compared to prior systems. This is an extremely important capability for any useful world simulator—you must be able to model failure, not just success.

The model is also a big leap forward in controllability, able to follow intricate instructions spanning multiple shots while accurately persisting world state. It excels at realistic, cinematic, and anime styles.

Prompt: intense anime battle between a boy with a sword made of blue fire and an evil demon demon

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

Prompt: Two mountain explorers in bright technical shells, ice crusted faces, eyes narrowed with urgency shout in the snow, one at a time

You can also directly inject elements of the real world into Sora 2. For example, by observing a video of one of our teammates, the model can insert them into any Sora-generated environment with an accurate portrayal of appearance and voice. This capability is very general, and works for any human, animal or object.

Prompt: Bigfoot is really kind to him, a little too kind, like oddly kind. Bigfoot wants to hang out but he he wants to hang too much

The model is far from perfect and makes plenty of mistakes, but it is validation that further scaling up neural networks on video data will bring us closer to simulating reality.

Deployment of Sora 2

On the road to general-purpose simulation and AI systems that can function in the physical world, we think people can have a lot of fun with the models we’re building along the way.

We first started playing with this “upload yourself” feature several months ago on the Sora team, and we all had a blast with it. It kind of felt like a natural evolution of communication—from text messages to emojis to voice notes to this.

So today, we’re launching a new social iOS app just called “Sora,” powered by Sora 2. Inside the app, you can create, remix each other’s generations, discover new videos in a customizable Sora feed, and bring yourself or your friends in via cameos. With cameos, you can drop yourself straight into any Sora scene with remarkable fidelity after a short one-time video-and-audio recording in the app to verify your identity and capture your likeness.

Last week, we launched the app internally to all of OpenAI. We’ve already heard from our colleagues that they’re making new friends at the company because of the feature. We think a social app built around this “cameos” feature is the best way to experience the magic of Sora 2.

Launching responsibly

Concerns about doomscrolling, addiction, isolation, and RL-sloptimized feeds are top of mind—here is what we are doing about it.

We are giving users the tools and optionality to be in control of what they see on the feed. Using OpenAI's existing large language models, we have developed a new class of recommender algorithms that can be instructed through natural language. We also have built-in mechanisms to periodically poll users on their wellbeing and proactively give them the option to adjust their feed.

By default, we show you content heavily biased towards people you follow or interact with, and prioritize videos that the model thinks you’re most likely to use as inspiration for your own creations. We are not optimizing for time spent in feed, and we explicitly designed the app to maximize creation, not consumption. You can find more details in our Feed Philosophy⁠

This app is made to be used with your friends. Overwhelming feedback from testers is that cameos are what make this feel different and fun to use—you have to try it to really get it, but it is a new and unique way to communicate with people. We’re rolling this out as an invite-based app to make sure you come in with your friends. At a time when all major platforms are moving away from the social graph, we think cameos will reinforce community.

Protecting the wellbeing of teens is important to us. We are putting in default limits on how many generations teens can see per day in the feed, and we’re also rolling out with stricter permissions on cameos for this group. In addition to our automated safety stacks, we are scaling up teams of human moderators to quickly review cases of bullying if they arise. We are launching with Sora parental controls⁠ via ChatGPT so parents can override infinite scroll limits, turn off algorithm personalization, as well as manage direct message settings.

With cameos, you are in control of your likeness end-to-end with Sora. Only you decide who can use your cameo, and you can revoke access or remove any video that includes it at any time. Videos containing cameos of you, including drafts created by other people, are viewable by you at any time.

There are a lot of safety topics we’ve tackled with this app—consent around use of likeness, provenance, preventing the generation of harmful content, and much more. See our Sora 2 Safety doc⁠ for more details.

A lot of problems with other apps stem from the monetization model incentivizing decisions that are at odds with user wellbeing. Transparently, our only current plan is to eventually give users the option to pay some amount to generate an extra video if there’s too much demand relative to available compute. As the app evolves, we will openly communicate any changes in our approach here, while continuing to keep user wellbeing as our main goal.

We’re at the beginning of this journey, but with all of the powerful ways to create and remix content with Sora 2, we see this as the beginning of a completely new era for co-creative experiences. We’re optimistic that this will be a healthier platform for entertainment and creativity compared to what is available right now. We hope you have a good time :)

Sora 2 availability and what’s next

The Sora iOS app⁠ is available to download now. You can sign up in-app for a push notification when access opens for your account. We’re starting the initial rollout in the U.S. and Canada today with the intent to quickly expand to additional countries. After you’ve received an invite, you’ll also be able to access Sora 2 through sora.com⁠. Sora 2 will initially be available for free, with generous limits to start so people can freely explore its capabilities, though these are still subject to compute constraints. ChatGPT Pro users will also be able to use our experimental, higher quality Sora 2 Pro model on sora.com⁠ (and soon in the Sora app as well). We also plan to release Sora 2 in the API. Sora 1 Turbo will remain available, and everything you’ve created will continue to live in your sora.com⁠ library.

Video models are getting very good, very quickly. General-purpose world simulators and robotic agents will fundamentally reshape society and accelerate the arc of human progress. Sora 2 represents significant progress towards that goal. In keeping with OpenAI’s mission, it is important that humanity benefits from these models as they are developed. We think Sora is going to bring a lot of joy, creativity and connection to the world.

— Written by the Sora Team

Sora 2

Generated by RSStT. The copyright belongs to the original author.

Source