ZeyuLing/Delighting-Benchmark-Diverse-v1

Paint3D vs Hunyuan3D-2.0 Diverse De-lighting Benchmark

This repository contains a category-balanced benchmark for comparing how Paint3D and Hunyuan3D-2.0 handle image-conditioned 3D texturing when the input image contains visible illumination and shading.

Contents

• —50 Google Scanned Objects cases from 10 categories (5 cases per category).
• —Strongly lit conditioning images and reference unlit albedo renders.
• —100 canonical method packages: 50 Paint3D and 50 Hunyuan3D-2.0 outputs.
• —Two 360-degree GT videos per case: one with source lighting/shadows and one unlit.
• —Per-case comparison images/videos, category contact sheets/videos, metrics, and an HTML browser.

The categories are footwear, bags_apparel, electronics, tools_office, kitchen_appliances, toys_animals, toy_transport_building, food_packaging, home_storage_decor, and games_puzzles.

Repository layout

benchmark_inputs/
  samples_all.txt
  manifest.jsonl
  input_contact_sheet.jpg
  <case>/
    input.png
    reference_albedo.png
    lighting_delta.png
    model.obj
    metadata.json
    source/
canonical_outputs/
  paint3d/<case>/
  hunyuan3d/<case>/
visualization/
  index.html
  metrics.csv
  category_summary.csv
  <case>/comparison.png
  <case>/comparison.mp4

Every directory under canonical_outputs/{paint3d,hunyuan3d}/<case> contains exactly seven files:

input_image.png
input_mesh.obj
textured.glb
albedo_render.mp4
albedo_texture.png
gt_lit_render.mp4
gt_unlit_render.mp4

The method comparison videos are H.264/yuv420p, 3072 x 1024, 25 FPS, and 100 frames. The two GT turntables are H.264/yuv420p, 768 x 768, 25 FPS, 100 frames, and 4 seconds. All 100 canonical packages and all 200 GT file placements passed file-contract and decoded-video validation. Reports are stored at canonical_outputs/canonical_packaging_validation.json, canonical_outputs/gt_validation_report.json, and canonical_outputs/gt_placement_validation.json.

Summary metrics

Lower is better. Across all 50 cases:

These metrics are proxy measurements and remain sensitive to geometry and view alignment. The comparison videos and contact sheets should be used for qualitative assessment of baked-in lighting and shadow removal.

Source and license

The source assets come from Google Scanned Objects:

• —https://research.google/pubs/google-scanned-objects-a-high-quality-dataset-of-3d-scanned-household-items/
• —https://github.com/kevinzakka/mujocoscannedobjects

Google Scanned Objects is distributed under CC BY 4.0. Source attribution and asset identifiers are retained in each case's metadata.json.