Bernice · Virtual human case study

5 minute read · 2023–24

What happens whenwe all converge onthe same face?

So I built one. Bernice is an AI virtual human I designed and built to live inside the logic of algorithmic beauty — and talk back to it.

Role: Concept, build & narrative — solo
Context: Individual M.S. thesis · Columbia GSAPP · 2023–24
Built with: Unreal Engine · MetaHuman · AR facial capture · Midjourney

Bernice — a MetaHuman render, a face engineered toward one idea of perfect — Bernice — a MetaHuman render. A synthetic face, engineered toward a single idea of “perfect.”

01 / Observation

Social platforms became a school of taste.

“Looks AI-generated” had quietly become a compliment.

Scroll a beauty or lifestyle feed long enough and the faces start to rhyme — the same jawline, the same “A4 waist,” the same dewy skin, ranked and amplified until sameness reads as aspiration.

To see it for myself, I collected beauty posts by hashtag and looked at the sheer attention behind each “ideal,” then overlaid creator headshots to trace how narrow the accepted range had become. It was a hand-collected, qualitative read — not a rigorous dataset — but the convergence was hard to miss.

#become beautiful#变美133.3B

#daily outfits#每日穿搭90.1B

#looks#颜值66B

#outfit#穿搭49.1B

#body#身材40.5B

#pretty girls#美女17.4B

Beauty hashtags on Xiaohongshu, by public view count (billions). A curated sample, not a full dataset — the scale of the convergence was the point.

Front-view overlay study: rule-of-thirds, golden-ratio and feature-width lines over a face blurred beyond recognition — The “ideal,” measured. My overlay study — creator headshots blurred beyond recognition, with only the measurement lines kept.

02 / The question

So I built the logical endpoint.

If convergence has a destination, it looks like a synthetic person — more “perfect,” more controllable, and more tireless than any human creator. Rather than write about it, I decided to make one, and see what building it would teach me.

The thesis

Bernice: a virtual influencer built to live inside the logic of algorithmic beauty — and talk back to it.

Then I put her where we’d actually meet her — composited into a real New York street. Almost passing; the almost is the point.

03 / The build

One person, the whole pipeline.

I’m not an engineer; I directed the tools. I built Bernice as a MetaHuman in Unreal Engine, drove her performance with iPhone AR facial capture, and used Midjourney for her imagery and ChatGPT to draft her posts — then gave her a voice, as a first-person film and a virtual account.

There’s an irony I leaned into: I couldn’t hand my measurements to the software. Bernice’s face came from MetaHuman’s built-in presets — a handful of “defaults” that already encode an idea of beauty before you touch a slider. The tool proved the thesis better than my analysis could, and building it end to end is how I found that out.

Model

MetaHuman in Unreal Engine

Perform

iPhone AR facial capture, Live Link

Generate

Midjourney imagery · ChatGPT scripts

Publish

A first-person film and a virtual account

Driving Bernice live: my own expressions on the phone (Live Link Face), her MetaHuman on the green screen in Unreal.

04 / The character

A virtual influencer, self-aware on purpose.

“From the relentless pursuit of being perfect to a journey of release and acceptance.”

I wrote Bernice a voice — a first-person film and a nine-post narrative arc that moves from the pressure to be flawless toward release and self-acceptance. I designed all nine; I published two before hitting the wall that became the project’s sharpest finding.

In 2024, the tools couldn’t hold her identity steady. Midjourney redrew a subtly different face each time, drifting off Bernice unless I retrained the model — so a “consistent” synthetic influencer simply wasn’t operable yet. The account and its follower count are a constructed thesis prop, not a real audience; what’s real is what the attempt taught me about where the technology broke.

Bernice's published Instagram post, a Midjourney-generated portrait in a snowy cabin

A second frame from the same published Instagram post, with a visibly different face

The two posts that actually shipped (@bbernice2042). Look closely: the face drifts between frames — the consistency limit, visible in the published work itself.

05 / Virtual vs. real

What you gain, and what you give up.

I mapped synthetic creators against human ones across the tradeoffs a platform or brand would actually weigh. The pattern underneath: the most controllable creators are the ones people trust least — the tension synthetic media keeps walking into.

Tradeoff	Synthetic influencer	Human creator
Image control	Total — every pixel is a decision	Bounded by a real body and life
Output	Continuous; never rests or ages unless scripted	Finite; needs time and rest
Cross-cultural reach	Re-skinned per market on demand	Rooted in one identity
Controversy risk	Low; every word is authored	Higher; unscripted and human
Authenticity	Manufactured	Lived
Audience trust	Fragile the moment it is disclosed	Earned, and durable

06 / Two years on

The question moved faster than the tools.

What it was

A critical build: a virtual human, a pipeline I ran end to end as a non-engineer, and a thesis about what algorithmic feeds do to how we want to look.

What I’d do now

Character consistency was the wall in 2024; the models and tooling since have largely dissolved it. I’d treat the evidence systematically instead of a hand-picked wall, and push the build toward something interactive rather than a fixed narrative.

What I keep from it
Build the thing yourself, early — so you judge the technology from the inside, including what it does to the people who’ll use it.

In one line

A face that almost passes — and the almost is the point.

Back to selected work