Google Makes It Easy to Deepfake Yourself

谷歌让“深度伪造”你自己变得轻而易举

A wave of déjà vu washes over me as Elias Roman, vice president of product management at Google Labs, demos a new “avatar” feature for Flow, the company’s tool that lets users generate and remix AI videos and images. He previously scanned his likeness to create a digital clone of himself. Now, he can insert himself into any AI-generated videoclip he wants using Google’s new Omni Flash model. 当谷歌实验室产品管理副总裁 Elias Roman 演示 Flow 的一项新“头像”（avatar）功能时，一种似曾相识的感觉涌上心头。Flow 是谷歌推出的一款工具，允许用户生成和重混 AI 视频及图像。Roman 此前已经扫描了自己的形象，创建了一个数字克隆体。现在，他可以使用谷歌全新的 Omni Flash 模型，将自己插入到任何他想要的 AI 生成视频片段中。

“This is for creators who want to bring themselves into their content but don’t want to have to shoot themselves,” Roman says. “这是为那些想把自己融入内容，却又不想亲自拍摄的创作者准备的，”Roman 说道。

This specific style of social-first, selfie deepfake is reminiscent of a quintessential feature from OpenAI’s now-defunct Sora app—rather than cameos or characters, Google calls them avatars. These avatars are also available through the Gemini app and YouTube. Google announced the new feature at its annual I/O developer conference in Mountain View, California. 这种以社交为先、自拍式的深度伪造风格，让人想起 OpenAI 现已停用的 Sora 应用中的一项典型功能——谷歌将其称为“头像”，而不是客串角色或虚拟人物。这些头像也可以通过 Gemini 应用和 YouTube 使用。谷歌在加利福尼亚州山景城举行的年度 I/O 开发者大会上宣布了这一新功能。

Google launched Flow last year under its experimental Labs division. “Google has never had a product line for creative work before,” Roman says. “Productivity, definitely. Developers, absolutely. Video consumption, yes. Not for creative work.” He sees this as Google’s attempt to build tools for the next generation of creators. 谷歌去年在其实验性实验室部门下推出了 Flow。“谷歌以前从未有过针对创意工作的产品线，”Roman 说，“生产力工具肯定有，开发者工具绝对有，视频消费产品也有，但就是没有针对创意工作的。”他认为这是谷歌为下一代创作者打造工具的一次尝试。

Similar to other announcements from Google I/O surrounding Google Search, many of the new changes to Flow are part of the company’s larger attempt to make AI agents, essentially automated software taskmasters, and vibe coding, building bespoke features with natural language prompts to AI, more mainstream for a broader audience. For example, users can repeat custom instructions when generating videos and create automated workflows that sort similarly styled clips into folders. 与谷歌 I/O 大会上围绕谷歌搜索的其他公告类似，Flow 的许多新变化是该公司更大规模尝试的一部分，旨在让 AI 代理（本质上是自动化的软件任务主管）和“氛围编程”（vibe coding，即通过自然语言提示向 AI 构建定制功能）为更广泛的受众所接受。例如，用户在生成视频时可以重复自定义指令，并创建自动工作流，将风格相似的片段分类到文件夹中。

One of the most immediately noticeable changes to Flow is the new video-generation model powering the experience: Omni Flash, succeeding Veo. Similar to how Google’s Nano Banana model brought more context about the world into the AI image-creation process, the Omni Flash model overhauls video generation with richer detail throughout clips. Flow 最直观的变化之一是驱动该体验的新视频生成模型：Omni Flash，它取代了之前的 Veo。正如谷歌的 Nano Banana 模型为 AI 图像创作过程引入了更多世界背景信息一样，Omni Flash 模型通过在整个片段中提供更丰富的细节，彻底革新了视频生成效果。

Flow users can generate characters in AI videos with more consistency via the Omni Flash model. Roman says this is a major improvement over the weakness in past versions of Flow, where created characters could warp during successive video generations. Also, a key character that Flow users can now generate in an AI scene after an AI scene? Themselves. 通过 Omni Flash 模型，Flow 用户可以在 AI 视频中更连贯地生成角色。Roman 表示，这相比 Flow 过去版本是一个重大改进，因为过去版本中创建的角色在连续生成视频时可能会变形。此外，Flow 用户现在可以在一个又一个 AI 场景中生成的关键角色是谁？正是他们自己。

Users set up an “avatar” of themselves by going into the settings of their Flow account and scanning a QR code on their phone. Then, Google asks users to record themselves saying a string of numbers aloud and move their head around to capture every angle. This selfie-capture style will feel familiar to anyone who signed up for the Sora app, which OpenAI launched last year as an AI-first social media platform where people can generate and share clips of themselves. OpenAI startlingly wound it down after less than seven months. 用户只需进入 Flow 账户设置并扫描手机上的二维码，即可设置自己的“头像”。随后，谷歌会要求用户录制一段朗读数字的视频，并转动头部以捕捉各个角度。这种自拍捕捉方式对于任何注册过 Sora 应用的人来说都会感到熟悉。OpenAI 去年推出了 Sora，将其作为首个 AI 社交媒体平台，人们可以在上面生成并分享自己的视频片段。令人惊讶的是，OpenAI 在不到七个月后就将其关闭了。

Unlike the Sora app, where users could generate videos of other users depending on the person’s settings, Google’s initial focus with its avatars is to let users create AI versions of themselves only, not other people. Every video generated with the Omni model, including those with your avatar, includes Google’s SynthID watermark. 与 Sora 应用不同（Sora 允许用户根据对方的设置生成其他用户的视频），谷歌目前对头像功能的重点是让用户仅能创建自己的 AI 版本，而非他人。使用 Omni 模型生成的每一段视频，包括包含你头像的视频，都带有谷歌的 SynthID 水印。

“You can capture your voice and your visual identity from multiple angles and have that show up with pretty high levels of fidelity,” Roman says. He generated a tongue-in-cheek video of himself teasing the Flow team in front of a dumpster fire, with an AI version of himself that looked lifelike and sounded like him. Then he used Flow to request changes to the generation, like a new background setting and a different-colored shirt, and Omni Flash adjusted the clips while preserving the avatar’s details. “你可以从多个角度捕捉你的声音和视觉形象，并以相当高的保真度呈现出来，”Roman 说。他生成了一段自嘲视频，视频中他在垃圾桶火堆前调侃 Flow 团队，AI 版的他看起来栩栩如生，声音也与他本人无异。随后，他使用 Flow 对生成内容提出修改要求，例如更换背景设置和改变衬衫颜色，Omni Flash 在保留头像细节的同时调整了视频片段。

This isn’t the first time Google has rolled out a version of self-controlled deepfake video tools for creators—last month, YouTube Shorts added a limited option for users to make similar AI avatars that can be inserted into clips on that platform. Other Silicon Valley companies are also looking for ways to transform creators’ outputs using generative AI. For example, last year, Meta rolled out an AI feature that can seamlessly translate Instagram Reels into different languages, even adjusting creators’ lips to match the different voices. 这并不是谷歌第一次为创作者推出自控式深度伪造视频工具——上个月，YouTube Shorts 增加了一项有限的功能，允许用户制作类似的 AI 头像并插入到该平台的视频片段中。其他硅谷公司也在寻找利用生成式 AI 改变创作者产出的方法。例如，去年 Meta 推出了一项 AI 功能，可以无缝地将 Instagram Reels 翻译成不同语言，甚至能调整创作者的唇形以匹配不同的语音。

While these AI tools may streamline aspects of the content production pipeline for creators—you don’t even have to get out of bed now to generate sassy vertical videos—generative AI is increasingly polarizing audiences who see these videos as inauthentic or not aligned with their values. Well, that’s if they actually clock the videos as AI. 虽然这些 AI 工具可能会简化创作者的内容生产流程——你现在甚至不需要起床就能生成时髦的竖屏视频——但生成式 AI 正日益让受众产生分歧，他们认为这些视频不够真实，或者与他们的价值观不符。当然，前提是他们能识别出这些视频是 AI 生成的。