PhysFlow: Frequency Decoupled with Dual-Field Rectified Flow for Remote Photoplethysmography

Overview

Remote Photoplethysmography (rPPG) is a non-invasive technique that estimates vital signs like heart rate by detecting subtle color variations in facial videos. However, these signals are often obscured by external disturbances such as illumination changes, head motion, and facial expressions. Existing deep learning methods often struggle to separate these weak physiological signals from strong noise, leading to unstable waveform reconstruction.

The PhysFlow Approach

PhysFlow addresses this by moving away from unified signal modeling. The authors propose a frequency-decoupled approach that splits the ground-truth rPPG signal into two distinct components: a trend component (representing the slow-moving waveform profile) and an amplitude component (capturing the fine-grained pulse variations).

Instead of treating these as auxiliary information, PhysFlow uses them as explicit supervisory targets. The model learns two separate conditional velocity fields—one for the trend and one for the amplitude—using a rectified flow formulation. This design reduces interference between the components, as each is modeled under its own objective. By leveraging the velocity-matching properties of rectified flow, the framework can reconstruct the rPPG signal efficiently using only a few ordinary differential equation (ODE) integration steps, making it suitable for real-time applications.

By explicitly separating the signal components, PhysFlow achieves higher robustness against the common environmental disturbances that plague contactless health monitoring. The use of rectified flow provides a significant computational advantage over traditional diffusion-based models, which typically require many iterative denoising steps. This work demonstrates that focusing on the morphological fidelity of the rPPG waveform, rather than just heart-rate accuracy, leads to more reliable physiological sensing in unconstrained, real-world scenarios.