Sahana Vijai Art Lead · Creative Technologist

I direct the visual work, and I build the systems that ship it. Two jobs that usually sit on either side of a handoff.

Meta

2015 – 2022 · Research group

Facebook Reality Labs Codec Avatars

One person, photoreal.
Do they accept it?

2022 – 2026 · Production team

Meta Horizon Avatars 2.0

Everyone, stylized.
Do billions accept it?

Will people accept this as themselves? Same question, opposite ends of the range.
Meta Horizon Avatars 2.0

I led hair as a category through the Avatars 2.0 launch.

Meta Avatars 2.0 is Meta's new stylized avatar system: how people show up across its apps. I ran the hair category end to end through the 2024 launch, covering the look, the catalog, the production system behind it, and the team that built it.

Meta Avatars 2.0 key art: stylized avatars mid-action, showing a range of hair types and styles.
Avatars 2.0, live. The result is public → meta.com/avatars
~100 hairstyles
shipped in the new style, still the live catalog today
0 → 90%
of the legacy catalog covered, in one half, so existing users kept their look
5 surfaces
Facebook, Instagram, WhatsApp, Messenger and Horizon, available to billions
The eye

Visual direction for hair inside the art director's style standard: a curated launch catalog, a 3D style guide for the category, and the representation framework that defined what coverage had to mean.

The hand

The production system underneath it: pipeline, vendor onboarding, quality rubrics and batch review, plus the hair-sim feasibility bridge between vis-dev and tech art.

Rebuilding a catalog looks like a scoping problem until you ask what it actually has to cover. Defining that, and making it measurable, is what turned representation into something the team could plan and fund rather than assess at review.

How we defined coverage →
Facebook Reality Labs Codec Avatars

I found out what people would accept as themselves.

Codec Avatars is Meta's photoreal avatar research: rebuilding one specific person closely enough that the reconstruction reads as them. One person at a time, each of them someone you could ask. I was on it for six and a half years, at Facebook Reality Labs in Pittsburgh.

The eye

Knowing a smile is wrong before you can say why. Then making the right one, because “wrong” is not something a system can be aimed at.

The hand

Handheld scanning, scan processing, sculpting, expression shapes authored on the facial rig, and cloth simulation. A new tool each time the research moved.

Everything we learned came from a small group of people we already knew. The work was deciding which of those findings would still hold for someone we would never meet.

What transferred, and what didn't →
Craft

The eye, trained on a decade of characters.

The eye

Knowing why a silhouette reads, why a face holds. Taste I earned over years of character work.

The hand

The topology, the sculpt, the technical craft that turns the read into a built asset.

Concept sculpt Expression Study The roar isn't a pose. It's a specific set of FACS shapes, sculpted to test whether human facial tracking could still drive a face built this far from a human skull.
Concept sculpt Orc An anatomy study aimed at age rather than power. Slack skin, asymmetry and a broken tusk are far less forgiving than muscle, and they do more for character.
Production-ready Goblin Built from a flat concept image, so the job was fidelity, not invention. Everything the drawing never showed had to be inferred without contradicting it. Taken all the way through: retopologised, UV'd, textured, rigged.
Production-ready Nyx Four material behaviours in one asset: hard-surface plate, soft jacket, skin, high-gloss latex. Each wants different topology and texel density; making them share one character is the real problem.
Expression Study, view 1 of 7Expression Study, view 2 of 7Expression Study, view 3 of 7Expression Study, view 4 of 7Expression Study, view 5 of 7Expression Study, view 6 of 7Expression Study, view 7 of 7
Expression Study
Orc, view 1 of 5Orc, view 2 of 5Orc, view 3 of 5Orc, view 4 of 5Orc, view 5 of 5
Orc
Goblin, view 1 of 8Goblin, view 2 of 8Goblin, view 3 of 8Goblin, view 4 of 8Goblin, view 5 of 8Goblin, view 6 of 8Goblin, view 7 of 8Goblin, view 8 of 8
Goblin
Nyx, view 1 of 13Nyx, view 2 of 13Nyx, view 3 of 13Nyx, view 4 of 13Nyx, view 5 of 13Nyx, view 6 of 13Nyx, view 7 of 13Nyx, view 8 of 13Nyx, view 9 of 13Nyx, view 10 of 13Nyx, view 11 of 13Nyx, view 12 of 13Nyx, view 13 of 13
Nyx

Deep archive → ArtStation

Experiments

I build with AI to find out what I can direct.

Generating good work got cheap. Directing it exactly where you want it did not. The question I keep coming back to is not how much a model can do alone, but where you get to step back in, and whether what comes out is still yours.

The hand

The agentic systems, the tooling, the procedural pipelines. Built to find where the control points are.

The eye

Art direction applied to a new material. Making sure my own aesthetic comes through the model.

Codec Avatars, then Avatars 2.0, now agentic AI. The same job each time: finding out what a new technology can actually deliver, and what a person will accept from it.