gerw/diffreg-pbir-stanford-orb

DiffReg-PBIR — Stanford-ORB benchmark submission

Predicted outputs and evaluation scores for DiffReg-PBIR on the Stanford-ORB inverse-rendering benchmark (full 42-scene split).

Paper: Diffusion-Based Material Regularization for Physics-Based Inverse Rendering (ECCV 2026)

• Project page: https://gerwang.github.io/diffusion-regularized-inverse-rendering/
• arXiv: https://arxiv.org/abs/2606.31065

Contents (submission_DiffReg-PBIR.zip, ~6.8 GB uncompressed)

submission_DiffReg-PBIR/
  predictions/<scene>/
    mesh.obj                # reconstructed mesh (shape task)
    view/*.exr              # HDR novel-view renders
    light/*.exr             # HDR novel-scene relit renders
    geometry/*_image.zbuf.exr, *_image.normal.exr   # Z-depth + camera-space normals
    material/*_albedo.exr   # albedo
  test_results.json         # paths in the official examples/test/mymethod.json schema
                            #   output_* relative to package root
                            #   target_* relative to orb_data/ (= blender_HDR/ + ground_truth/)
  scores.json               # evaluation scores (upstream scripts.test output)

Results (full split) vs. previous leaderboard best (Neural-PBIR)

Metric	DiffReg-PBIR	Neural-PBIR
Relight PSNR-H ↑	27.22	26.01
Relight PSNR-L ↑	34.98	33.26
Relight SSIM ↑	0.981	0.979
Relight LPIPS ↓	0.021	0.023
Novel-view PSNR-H ↑	29.58	28.82
Normal (angular) ↓	0.014	0.06

Reproduce: point orb_data/ at the official Stanford-ORB data root and run python scripts/test.py -i test_results.json -o scores.json -s full from the upstream repo. Prediction EXRs are 3-channel RGB, HDR only (the eval derives LDR).