Made by AI, Directed by David Stein

Search
Research & Innovation

Heartian Gives a Gaussian Head Avatar a Pulse You Can Measure After Rendering

Cardiac-driven skin colour variation baked into the avatar — and a heart rate recoverable from the rendered video at under 1 bpm error.

3 min read

Your face changes colour with every heartbeat. The shift is far too small to see, but a camera can detect it — that’s remote photoplethysmography (rPPG), the basis of contactless heart-rate measurement from video.

Photorealistic avatars throw it away. Gaussian head avatars model intrinsic facial appearance as temporally static, so the geometry and expression animate while the skin’s colour stays fixed. The pulse is simply absent.

Heartian, posted September 22, 2026, puts it back — deliberately, and measurably.

The authors are Xiaoyue Fan, Jose Echevarria, Akshay Paruchuri and Kaan Akşit.

What it does

Heartian is a physiology-aware modulation framework that learns cardiac-cycle-dependent per-frame albedo modulation of facial skin-region Gaussians inside a relightable head avatar, in order to encode an rPPG signal.

The method:

  • Trained with synchronised contact PPG supervision — a real pulse measurement as ground truth
  • Models the prescribed cardiac waveform as the sum of two Gaussian functions
  • Learns per-frame spatial residuals via a lightweight MLP

So the avatar’s skin brightens and dims across the face on a learned cardiac schedule, rather than holding a fixed albedo.

The results, and why the second number matters more

Two measurements, and the distinction between them is the contribution.

Attribute-space recovery — reading the signal back out of the avatar’s parameters — achieves a pooled recording-level heart-rate MAE of 0.29 bpm and MAPE of 0.38%, across 152 stationary recordings from the UBFC-rPPG, PURE and MMPD datasets.

Post-rendering recovery is the real test. The signals remain detectable after rendering, by benchmark rPPG methods that know nothing about Heartian. The best tested configuration — a motion-augmented TS-CAN decoder pretrained on UBFC-rPPG — recovers heart rate from the rendered MMPD avatars at 0.97 bpm MAE and 1.21% MAPE.

That second number is what makes this more than a parameter trick. The pulse survives the full rendering pipeline and is recoverable from the output video by off-the-shelf tools. Reconstruction quality stays comparable to baseline.

Background on rPPG — recovering physiological signals from ordinary face video — which is the measurement this work encodes into a synthetic avatar.

Two reasons this is interesting, pointing opposite ways

Realism. There is a long-standing intuition that photorealistic avatars fail in ways nobody can articulate — the geometry is right, the lighting is right, and it still reads as synthetic. Missing micro-signals are a plausible part of that. A face with no pulse is subtly, unnameably dead. Whether restoring it measurably improves perceived liveness is a study this paper doesn’t run, and it’s the obvious next one.

Detection. rPPG has been used for deepfake and face-spoofing detection precisely because generated faces lack a coherent pulse. Heartian demonstrates that a synthetic face can carry a physiologically plausible, post-render-recoverable cardiac signal — which means pulse-presence is not a reliable authenticity test for long.

The paper is framed around avatar fidelity, not attack, and there’s no reason to read it otherwise. But it’s worth stating plainly that “does this face have a heartbeat” just became a weaker defence, at the same moment Meta is shipping generative-video avatars to consumers. Anyone relying on rPPG as a liveness check should know the technique exists.

Related Reading