Hey Gen, is that really you?
By now we’ve all seen examples of the Ai video avatars that the tech bro gods tell us are to be integrated into our future professional personas. It’s a deep fake future in which we’re being lead to accept never knowing if we’re hearing from or speaking to a real live human in our Zoom meetings, YouTube videos or even blockbuster movies. The Ryan Reynolds cast in “Free Guy 2” could be a legit NPC and maybe we won’t even know.
These are the kinds of ideas that come from the same sort of minds that want to build robot girlfriends out of Scarlett Johansson. And much like robot girlfriends, most people aren’t going to buy into your Ai video clone.
The company leading this movement is HeyGen. Launched in mid-2022 by Joshua Xu coming out of SnapChat and Wayne Liang from ByteDance, the startup bros secured a $60M Series A funding round led by Benchmark, valuing the company at $500M at the tender age of two years old in mid-2024. HeyGen stock avatars quickly became the personas of LLM’s who used them to calmly predict that they would probably exterminate humanity in a few years, for the lulz, of course.
What happened to Julia?
Not many humans creating content for a public audience have bought into the HeyGeneration’s plan of digitally cloning themselves. But one who has is Julia McCoy.
After extricating herself from her upbringing in a cult by becoming an author telling her story and employing a team of writers doing corporate work, she started to get traction on YouTube early in 2024 by becoming one of the first and few female creators to focus on Ai. But once HeyGen-type tech became viable enough to cross the uncanny valley for enough viewers, she disappeared, leaving her digital clone to recite the lines generated by an LLM fine tuned to write in her style by digesting the content of her early videos.
The results of Julia’s transition from human to clone seem to be at parity, with some videos getting fewer than 10k views, while a few others have gotten over 200k. Scrutiny of them tends to reveal certain deflections in the eyes that indicate the limits of the deep fake. But clearly a lot of people don’t care or are listening more than watching.
But lost in the digitization process are her improvized asides present in her human videos. We don’t get the gleam in her eyes as she brags about firing her entire staff of writers to replace them with LLM’s, or her wink-wink endorsements of President Trump. It makes her videos very safe and contained, focused purely on the subject matter. And for her content and personal brand, that’s probably for the best.
Other use cases in which the HeyGen approach to Ai cloning are clearly winners include business applications. Corporate HR managers can get a gleam in their own eyes as they cancel the contracts of human-run video production companies that used to produce onboarding and training videos, in favor of their very own Ai spokesperson who can even rival Scarlett Johansson if they want. Busy execs can deliver corporate announcements that are pitch perfect while they personally deliver the good news to their boards of directors that they’ve been able to lay off 90% of their most expensive human developers. It’s a win win!
So what do audiences prefer over the fake McCoy?
There’s another stylistic approach to the same strategy that may be proving to yield far better results for public facing creators. Instead of attempting to fool their audiences into experiencing a deep fake version of their human personas as real, they take a more artistic approach. An example of an Ai YouTube channel that has grown exponentially faster and larger than Ms. McCoy’s is Ai Revolution.
Ai Revolution has grown an audience that is 200% larger,
with 33.6M views in the last month compared to McCoy’s 4.8M. Both channels cover many of the same stories in the Ai business silo, and both use Ai clones to address their audiences. So what are the key differentiators?
Ai Revolution’s avatar makes no attempt to be a deep fake. Rather, the avatar is, logicially for an Ai channel, a talking humanoid robot. And while this approach is not very compatible with HeyGen’s app, there are of course other options. One that is similar and better in some ways, is Runway ML’s Act One. And it’s this unique tool that my team uses to produce animated Ai avatars for clients like Lars Murray, a veteran of the NYC music biz, and host of the industry’s indispensible daily podcast, Unlimited Supply.
Why use video at all, especially for a podcast?
Look, being one of the richest bros on the planet doesn’t come cheap, so of course Zuck has a lot of pressure on him to raise the rates of paid media on his plethora of platforms. So the days of slapping a few bucks on an audiogram for the Zucks are pretty well gone. These days you need to shop for audience like you shop for veggies and go organic. And that means that your podcast audio files are basically useless outside of Spotify, Apple Podcasts etc. To leverage the organic media growth engine you need to have video. And the more competitive that game gets, the more you need to visually differentiate.
So for Lars’ new Unlimited Supply growth marketing plan, we looked at how to leverage Ai avatar based video content in a style and strategy that would fit with his established personal brand, resonate with his established audience, and get the attention of newbs scrolling through their iPhone feeds.
So as a reward for reading this far, here’s a brief breakdown of the workflow we used to create and implement an Ai avatar for Lars, the right way.
Step 01: Make some art, man.
To get the ball rolling, we worked with Lars’ team’s head of design to render his likeness in an artistic style that fit his brand and established visual aesthetics. After a few iterations and experimentations with animations we arrived at an image that had a somewhat retro, hand-done look.
Step 02: Create a trainer.
Next we took Lars’ new headshot art and employed our old, but still very useful friend, Mr. Green Screen. This gave our Ai tools the clearest possible visual direction for what we wanted and discourged them from dropping acid and hallucinating a bunch of useless artifacts in the background.
Step 03: Record the human.
You know all those boring podcast clips you see, showing the host dude giving a promo schpiel to his laptop? Yeah well, we did that too.
Step 04: Animate the trainer.
Where the Ai avatar magic started to happen was when we combined our image trainer with our laptop video using Runway’s Act One, which is the current frontier model for capturing and emulating the nuances of facial expressions, and transferring those into a stylized image. Act One allows for up to 30 second generations.
Step 05: Animate the backgrounds.
First we took the 1:1 (square) logo images from the Unlimited Supply podcast, with different colors for each day of the week and thematic focuses, removed the text logo/day labels, and reformatted them as 9:16 (vertical) images. Then we used Runway’s Gen3 Alpha model to create subtle motion animations for each. *Note that most other models we tried for this failed by going way over the top. Gen3 Alpha had a unique talent for subtlty.
Step 06: Comp the avatar & BG’s.
Once we had our animated avatar on a clean green and our nicely flowing animated background files sorted, we brought the project into our Ai-powered Swiss Army knife workhorse, Descript. The first thing Descript does when uploading spoken content is to transcribe, label, and make it editable via text UI, as well as the traditional timeline format used by the Avid-era apps.
Step 07: Optimize the audio.
Another essential function at which Descript excels is audio optimization. In addition to the usual tools like EQ, compressor, limiter etc, Descript offers an ai-powered noise gate branded, “Studio Sound,” which automates the process of removing background noise and normalizing the raw audio file. You know all those new features in Adobe CC? Yeah, that’s Adobe trying to catch up with Descript.
Step 08: Motion graphics & music.
Since Unlimited Supply is an established podcast focused on the music business then Lars already had a groovy synth theme ready to use. Otherwise we’d probably have to include a tutorial for Suno here. Leave a comment if you’re interested in learning more about Ai tools for generating music. For the motion graphics to animate onscreen transcript text and other labels, Descript includes a very useful set of tools, without trying to reinvent After Effects. So not an advanced specialty tool, but a fantastic utility that adds dynamism and clarity.
Bottom line on Ai Avatars
To wrap this up I’ll just reiterate that if you need internal corporate content for practical purposes like onboarding, training videos, or corporate announcements, then HeyGen, and competitors such as Synthesia, and Jogg, are very useful Ai tools. But if you’re trying to compete in the public content arena, you will need to work harder at it, and Runway’s Act One is the superior tool for ambitious creators. Using artfully rendered avatars will engage your established audience, attract new members to your audience, and help nurture them by letting them know that when they see your human face on their screens, they’re getting the real you, not a deep fake.