Stereon — Capture-Time Geometric Verification
A research direction into verifying image authenticity through physical-scene geometric consistency rather than learned appearance artifacts.
Jul 2026
Overview
Stereon is a research direction I'm pursuing, picking up where TruPhoto left off. TruPhoto established a ceiling: appearance-based detectors learn generator-specific artifacts, so they degrade against architectures designed to erase those artifacts. StyleGAN3 at 37% accuracy is the honest limit of that approach.
Stereon asks a different question. Instead of inspecting an image for traces of how it was made, it tests whether a set of images is geometrically consistent with a real physical scene. Multiple views of a real object triangulate to one coherent 3D structure; a photograph of a screen reconstructs as a plane; a set of generated views fails to agree on any single structure at all. The signal is physics rather than learned artifacts, so it generalizes by construction — it does not care which model produced the image.
This line of work continues in ProofShape, which applies the same geometric-consistency idea to inspecting manufactured parts.
Filing
USPTO provisional application 64/018,423, "System and Method for Multi-Signal Physical Scene Geometry Verification for Image Authenticity Detection," filed March 27, 2026. Sole inventor.