At the annual Google I/O developer conference, Elias Roman, Google Labs’ Vice President of Experimental Product Management, unveiled a new feature for the company’s Flow service: personalized avatar creation. Flow, a tool designed for users to generate and edit AI-powered videos and images, will now allow content creators to scan their likeness, produce an accurate digital clone, and integrate it into any AI-generated video clip. Roman explained that this innovation is intended for creators who wish to appear on camera without the time commitment of actual filming.
This method of generating personalized digital likenesses bears a resemblance to a key feature of OpenAI’s now-defunct Sora application. However, Google’s version officially refers to these digital copies as avatars and plans to make them accessible not only within Flow but also across the AI assistant Gemini’s ecosystem and the YouTube video platform. Elias Roman highlighted that Flow, launched last year, marks a significant milestone as Google’s first product suite designed not for coding, productivity, or content consumption, but purely for creative expression.
The integration of avatars into Flow aligns with Google’s broader strategy, as articulated at Google I/O. The corporation is committed to normalizing AI agents and the concept of “vibe coding” for a mainstream audience. For instance, Flow users can now establish recurring instructions for video generation and configure automated processes to sort clips of similar styles into designated folders.
The new Omni Flash video generation model, replacing the Veo neural network, is the primary technological driver behind the updated service. Much like the Nano Banana image generation neural network enhanced image creation with a deeper comprehension of real-world context, Omni Flash elevates video production by imbuing each frame with rich detail. Furthermore, this new model addresses a critical issue in previous Flow versions: the distortion and warping of character features during the generation of sequential scenes.
The process for creating a personal avatar involves users navigating to their Flow account settings and scanning a QR code with their smartphone. The system then prompts the user to record a brief video, reciting a sequence of numbers aloud and turning their head to capture their face from all angles. This workflow closely mirrors the functionality found in OpenAI’s Sora app, which was initially positioned as the first AI social network but unexpectedly ceased operations less than seven months after its launch. Crucially, Google imposes a strict guideline: only self-generated digital likenesses are permitted, not those of other individuals. Additionally, all videos featuring these avatars will be embedded with an invisible SynthID digital watermark.
As a demonstration, Elias Roman presented a humorous short video where his lifelike digital double, speaking in his actual voice, admonished the Flow team against the backdrop of a trash can. Roman then used on-the-fly text commands within the Flow interface to alter the video’s background and his avatar’s shirt color. The Omni Flash model instantaneously reconfigured the scene while preserving the avatar’s appearance.
This is not Google’s first foray into deepfake tools; a limited AI avatar creation feature was introduced to YouTube Shorts just last month. Competitors are also advancing similar technologies. For example, Meta* integrated an AI translator for Instagram* Reels last year, which not only modifies the dubbed language but also synchronizes the speaker’s lip movements with the spoken words.
For content creators, these new tools promise a significant simplification of production workflows. On the other hand, generative AI is increasingly polarizing audiences, who may perceive such content as artificial or inauthentic, especially if viewers can even discern that they are viewing a digital replica rather than a real person from the outset.
About the Author
