Goal
Test a hunch: that the per-pixel variance fields of a path-traced stereo pair are correlated once aligned by ground-truth disparity — even though the per-sample Monte Carlo noise streams are independent by construction — and that the correlation is strong enough to act as a learnable matching cue.
Progress
The correlation is real. Across 20 Mitsuba 3 scenes the warped Pearson correlation reaches ρ ≈ 0.754, invariant over a 16× sampling range, strongest in Lambertian regions and weaker in glass. A block-shuffle intervention showed the structure functions as a usable cue at the cost-volume level — a sim-to-real gap mechanism unique to rendered data.
That finding is written up and posted as a preprint: Independent Samples, Correlated Variance A Learnable Cross-View Cue in Path-Traced Stereo Data (arXiv v2) (arXiv, DOI 10.48550/arXiv.2606.25483; revised September 20, 2026).
Where it’s going
The preprint establishes that the shortcut exists. The open question — and the reason this stays in progress — is whether stereo networks actually exploit it in practice, and what that implies for training on rendered data. That’s the thread I’m still pulling.