Realtime avatars are too expensive to be everywhere — what if runtime cost was near zero?

This is exactly the question I asked myself when I looked into realtime avatars for my own projects.

The tech is getting impressive, but the economics still felt broken to me. Especially for B2C products or smaller B2B use cases, paying heavily for every live avatar minute makes a lot of ideas hard to justify.

So I started digging into the technical side and built PhotoCall, a pipeline that turns a single photo into a photorealistic realtime avatar.

The basic idea: move the expensive work into preprocessing instead of doing everything live at runtime.

That means the avatar can run with very low latency and very low runtime cost, while staying photorealistic and reproducible.
The tradeoff is lip sync precision. It won’t be as accurate as doing GPU-heavy frame-by-frame video generation at runtime, but for many product use cases that feels like a good trade: reliable, photorealistic characters that are instant, reproducible and cheap enough to use repeatedly.

The single-photo part also changes the user stories. If you don’t need a reference video, the same person or character can be reused across different scenes, looks, roles and contexts.

That feels like it could open up use cases beyond classic “talking head” demos: onboarding, product demos, support, interview prep, education, AI companions or something less obvious.

So I’m trying to figure out the right direction:

build this as infrastructure for other builders, or use it to build specific products on top?

What would you do?