How It Works

The physics and neural architectures behind Facial Authenticity Localization.

ICCV 2019 Optical Flow & ResNet Interpolation Artifacts

1. The Problem of Geometric Warping

Unlike synthetic media (such as diffusion models or GAN-generated deepfakes) that synthesize faces from noise, traditional photographic manipulation relies on geometric displacement. Tools like Adobe Photoshop's Face-Aware Liquify, Facetune, and smartphone beauty filters allow users to push, pull, stretch, and shrink facial features (e.g., thinning a jawline, reshaping a nose bridge, or enlarging eyes).

While visually seamless to the human eye, these operations alter the underlying grid of pixels through bicubic or bilinear resampling, leaving distinctive mathematical artifacts in the spatial frequency domain.

2. Facial Authenticity Localization (FAL) Architecture

Face Warp Check uses the methodology introduced by Wang et al. (ICCV 2019). The architecture is trained to perform two simultaneous tasks:

A. Global Classification: Calculates the probability that a given face image has undergone geometric warping.

B. Local Patch-Level Regression: Predicts an optical flow field $(\Delta x, \Delta y)$ across every spatial location, mapping each modified pixel back to its estimated original coordinate.

The pipeline consists of three core components:

  1. Face Alignment: Automatically identifies facial landmark anchors to crop and standardize the input image geometry.
  2. ResNet-50 Feature Extractor: Evaluates low-level interpolation fingerprints and high-level structural deformities.
  3. Flow Field Inversion (Undo): Uses the predicted vector field to reconstruct an approximation of the unmodified face.

3. Scoring Tiers & Interpretation

FAL outputs a continuous signal from 0.00 to 1.00. To ensure responsible and scientific communication, scores are mapped to descriptive tiers rather than binary "real/fake" judgments:

Signal Range Interpretation Label Clinical / Practical Significance
< 0.35 Low geometric-warp signal Resampling patterns match standard unmanipulated digital sensor capture.
0.35 – 0.70 Possible geometric warp Ambiguous warping artifacts or minor distortion consistent with gentle adjustments.
> 0.70 Strong geometric-warp signal Significant structural pixel displacements detected across facial landmark boundaries.

Note: The score measures the chance this face crop exhibits Face-Aware Liquify or similar warping characteristics. It is never a percentage probability that the entire photo is fraudulent.

4. Next Steps & Backend Roadmap

This web interface (v1) provides the structural layout and verification harness for the tool. Version 2 will interface with the PyTorch-based FAL inference service to generate real-time heatmaps and reverse warp vectors directly in the browser.