A parameter-free forecast of copper's changing color beat trained forecasting models and a last-frame copy on an unseen recording, according to the study. The method, called affinergb, improved on persistence, the practice of copying the last observed frame, by 13.4% when a chamber recording was used to forecast open-air copper and by 50.6% in the reverse direction. It was the only one of five entries to beat persistence in both tests, while every trained model performed worse than persistence.
The task used five observed frames at 128 by 128 pixels to predict the next 10 frames. Its main score was a mean-squared error calculated only on foreground pixels that actually changed, with the change threshold calibrated to the measured noise in each recording. The dataset comprised two fixed-camera copper recordings: one specimen aged in a controlled accelerated-corrosion chamber and one oxidizing in open air on a different day. For each transfer direction, 492 frames from the source recording were used for training, another 246 were reserved for checkpoint selection, and all 950 frames from the other recording made up the test set.
The advantage grew with the horizon
The pipeline forecasts 10 accelerated units ahead and turns the result into a material for a standard renderer. Its advantage over persistence grew with the forecast horizon. In the chamber-to-air direction, the relative improvement was 5.8% at t+1, 9.6% at t+5, and 16.7% at t+10. In the air-to-chamber direction, the figures were 19.8%, 47.4%, and 55.5% at the same horizons.
Why a simple forecast transferred
Affinergb is deliberately spare. It fits a global 3 by 3 color matrix and a bias from the first context frame to later frames, linearly extrapolates that sequence, and applies the resulting transform to the last observed frame. Here, global means that the same color transform is fitted for the image as a whole. The forecast therefore carries one image-wide color trajectory forward from the last frame.
The learned comparator, patinanet, takes a different route. It predicts a non-negative increment rather than the oxidation state itself, keeps the state non-decreasing for any forecast horizon, and uses a static susceptibility map to distribute change spatially. Checks on the source recordings indicated that the trained models optimized on the data they had seen. But on the unseen recording, every trained model was worse than simply copying the last frame.
Drift did not erase the gap
The researchers also checked whether measured drift in the recordings could account for the gap. They estimated change in a reference region and corrected every frame. In the chamber footage, the reference changed by 0.001 while the specimen changed by 0.128, a 105-fold difference. In the open-air footage, the reference brightened by 0.026 while the specimen changed by 0.119.
After that correction, the reported chamber-to-air margins were 17.3%, 19.9%, and 9.3%; the reverse-direction margins were 50.3%, 49.9%, 49.5%, and 46.7%. The correction did not remove the reported transfer advantage.
The uncertainty is more complicated than the headline result suggests. There were 810 overlapping test windows, but each 950-frame recording supplied only six independent 140-frame windows for the moving-block bootstrap, a resampling check. The interval for chamber-to-air was -13.9% to 51.8%, while the air-to-chamber interval was 35.6% to 62.7%. Only the air-to-chamber result was individually separated from zero, so the smaller chamber-to-air result was directionally consistent but not individually significant.
From pixels to a material
The output was not limited to a predicted image. The pipeline converted the forecast into albedo, normal, roughness, and metallic maps that a standard renderer could consume. On identical geometry and lighting, both forecasters beat a static canvas-only control at +5, +10, +20, and +50 accelerated units. At +10, affinergb had a FLIP score of 0.230 versus 0.453 for the control, and an LPIPS score of 0.163 versus 0.388. These are image-error measures in the reported comparison, so the lower scores indicate a closer match to the true future.
A narrow test with a useful warning
The warning comes from the size and design of the test. It used two copper recordings, with one held out in each transfer direction, and the recordings differed in both specimen and exposure condition: one came from the accelerated chamber and the other from open air on a different day. This means the result describes these two transfers, but cannot separate what came from the specimen from what came from the exposure condition or establish broad generalization across copper.
A map of where the model expected surface change to occur agreed with itself within a specimen at correlations of 0.63 to 0.85, but only at 0.34 across specimens. Compared with observed corrosion, it reached r = 0.68 on the training specimen and r = 0.00 on the held-out specimen. The authors treat this as a spatial pattern, not a specifically chemical one, because differences in pose and framing leave those explanations confounded.
The document is an arXiv preprint, version 1, dated 28 Aug 2026. The paper says its extraction scripts, leave-one-recording-out split builder, leakage audit, forecasters, renderer, and analysis are released, with tables and figures regenerable from run artifacts by a single command. The result remains a two-recording result.
Paper data and sources
Original title: What Will This Copper Look Like Later? Forecasting Surface Appearance and Rendering It as a PBR Material
Authors: Teejuta Sriwaranon, Borworntat Dendumrongkul, Tanapat Chamted, Pizzanu Kanongchaiyos
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text