Albedo Estimation using a Deep Restormer Architecture from Synthetic Data

Intrinsic albedo estimation is ill-posed: a single lit image confounds reflectance with shading, cast shadows, inter-reflections, and

tone mapping. Synthetic rendering supplies exact supervision, yet it is unclear how much capacity and spatial resolution a compact

transformer needs before it can exploit that supervision. We present a reproducible Blender/ModelNet10 pipeline that renders each

CAD mesh twice with identical geometry, camera, and base colour: a lit pass through a Principled BSDF with AgX tone mapping,

and an unlit albedo pass through an emission shader with a Standard transform. After integrity filtering, 5,234 lit–albedo pairs are

split 80/10/10. We adapt Restormer’s MDTA and GDFN blocks into a single-scale, full-resolution network and run a fair ablation

between a 44,445-parameter proof of concept (PoC: dim=24, 2+2 blocks, 1282

) and a 282,457-parameter production model (dim=48,

4+4 blocks, 2562

), both trained for the full 65 epochs. On the 524 held-out test images, the production Restormer reaches 32.22 dB

test PSNR, 0.0219 test L1 loss, and 0.9737 test SSIM. Granular row-by-row qualitative evaluation reveals nuanced visual trade-offs:

while the Production model decisively eliminates the severe boundary halos of the PoC on the sofa (Row 2) and sharpens fine filigree

on the chair (Row 3), both models fail to resolve the chromatic collapse on the table (Row 1), each erroneously predicting a light-blue

hue instead of green.