A High-Throughput Video Modem
We built an independent data channel carried inside an ordinary video call, invisible to H.264 compression — data riding along inside a completely normal-looking video stream.
Graphics & Synthetic Media Research
From invisible data in ordinary video to real-time synthetic faces.
Three lines of work, one shared subject: what actually happens to a video signal once it's compressed, and how much can be built on top of that — a hidden data channel, or a real-time synthetic performance — without breaking it.
- Two working data-hiding technologies
- A third built with no generative model
- Runs on ordinary CPU hardware
Data hidden in the picture itself
A channel that survives what compression is built to destroy.
Ordinary lossy video compression is designed to throw away anything a viewer won't notice — which is exactly why hiding a second signal inside it, so it survives that process, is a hard and interesting problem.
A Channel Built Into Compression Itself
A blind data channel built directly into the DCT math that H.264 compression itself is built on — engineered specifically to survive that compression rather than merely tolerate it. Developed with support from the Open Technology Fund.
Our newest technology
Real faces, not generated ones.
Facial Mimicry Cloning clones real facial motion from a short reference video — a few minutes is enough — and re-drives it from new text or audio, without any generative model. No GAN, no diffusion, no per-subject neural training.
Instead, real captured frames are indexed by their acoustic context and stored in a compact segment library. At playback, the method selects and smoothly blends real frames matching the incoming audio, using optical flow rather than generating new pixels — the face in the output was always a real, recorded face.
The practical result: no GPU dependency, no per-subject training step, and well under 5 seconds of end-to-end latency for a short utterance, running entirely on CPU. A separate silence-handling extension avoids the "frozen face" artifact during pauses, using a small state machine rather than an ad-hoc rule.
Where this connects
Why this matters beyond the lab.
A lightweight, present tutor
A CPU-only synthetic presence is cheap enough to give an AI tutor a real, attentive face without expensive infrastructure behind it — explored further in AI in Education.
Communication that keeps working
A data channel that survives ordinary video compression also survives the kind of network interference that blocks educational content and communication in some regions — the subject of Accessibility & Reach.
Curious about the technical details?
We're glad to talk to researchers and collaborators.