Research map / Video / Video segmentation and tracking
PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PhysDiff, a diffusion model for remote photoplethysmography that disentangles rPPG into trend and amplitude dynamics, denoises with a spatial-temporal hybrid transformer, and selects among multiple hypotheses by periodicity.
Problem
Remote PPG estimation from facial video maps frames directly to static rPPG values, ignoring the dynamic nature of the signal, and is fragile under illumination, head motion, and skin tone variation.
Approach
Apply a DDPM-style diffusion model to rPPG signals conditioned on MSTmap features and on disentangled trend and amplitude components of the noisy signal; add a disentanglement consistency loss and a PSD-based periodicity selection among multiple sampled hypotheses at inference.
Key results
- Intra-dataset on VIPL-HR: RMSE 6.15 and MAE improvement of 17.64% over NEST per the text; PhysDiff reaches 3.92 MAE and 6.65 RMSE on VIPL-HR in Table 1.
- Intra-dataset on UBFC-rPPG: MAE 0.33 and RMSE 0.57, improving RMSE by 10.93% over CPE.
- Cross-dataset UBFC-rPPG to MMSE-HR: RMSE 3.45 bpm, the lowest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the diffusion model (ablation) | PhysDiff · VIPL-HR | RMSE: 6.83 -> 6.37 | Helped |
| Add trend and amplitude disentanglement (ablation) | PhysDiff · VIPL-HR | RMSE: 6.37 -> 6.30 | Helped |
| Add the disentangled loss (ablation) | PhysDiff · VIPL-HR | RMSE: 6.30 -> 6.23 | Helped |
| Add periodicity-based multi-hypothesis selection (ablation) | PhysDiff · VIPL-HR | RMSE: 6.23 -> 6.15 | Helped |
| Use trend representation only (ablation) | PhysDiff · VIPL-HR | RMSE: 6.31 -> 6.26 | Helped |
| Use amplitude representation only (ablation) | PhysDiff · VIPL-HR | RMSE: 6.31 -> 6.30 | Helped |
| Use both trend and amplitude representations (ablation) | PhysDiff · VIPL-HR | RMSE: 6.31 -> 6.15 | Helped |
| Replace the hybrid denoiser with a stacked spatial-temporal Transformer (ablation) | PhysDiff · VIPL-HR | RMSE: 6.15 -> 7.01 | Hurt |
| Replace the hybrid denoiser with a parallel spatial-temporal Transformer (ablation) | PhysDiff · VIPL-HR | RMSE: 6.15 -> 6.78 | Hurt |
| Average hypotheses instead of selecting by periodicity (ablation) | PhysDiff · VIPL-HR | RMSE: 6.15 -> 6.19 | Hurt |
| Use PhysDiff rather than the strongest reported baseline | rPPG estimation · UBFC-rPPG | RMSE: 0.64 -> 0.57 | Helped |
| Use PhysDiff rather than the strongest reported baseline | rPPG estimation · PURE | RMSE: 0.98 -> 0.54 | Helped |
Limitations
- Cross-dataset UBFC-rPPG to PURE transfer still gives RMSE above 10 bpm, showing limited robustness to head movement.
- Inference uses multiple sampled hypotheses and iterative denoising, adding cost.
Experimental setup
Datasets: PURE UBFC-rPPG VIPL-HR MMSE-HR
Benchmarks: UBFC-rPPG PURE VIPL-HR MMSE-HR
Metrics: MAE RMSE Pearson correlation r
Baselines: DeepPhys PhysNet CVD PhysFormer NEST Contrast-Phys+ CPE Dual-GAN EfficientPhys TS-CAN
Builds on: DDPM DDIM PhysFormer MSTmap
Closest papers by meaning
- Remote Photoplethysmography in Real-World and Extreme Lighting Scenarios (CVPR 2025 · similarity 0.90)
- RhythmMamba: Fast, Lightweight, and Accurate Remote Physiological Measurement (AAAI 2025 · similarity 0.90)
- FactorizePhys: Matrix Factorization for Multidimensional Attention in Remote Physiological Sensing (NeurIPS 2024 · similarity 0.88)
- Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation (AAAI 2024 · similarity 0.88)
- FacePhys: State of the Heart Learning (NeurIPS 2026 · similarity 0.88)
- PaPaGei: Open Foundation Models for Optical Physiological Signals (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.