Reproduce “Appearance-Driven Automatic 3D Model Simplification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jon Hasselgren, Jacob Munkberg, Jaakko Lehtinen, Miika Aittala, Samuli Laine
An inverse-rendering framework automatically simplifies 3D assets by jointly optimizing mesh geometry, textures, materials, transparency, displacement, and skinning to match reference renderings across randomized views and lighting. It reduces complex assets to as little as 0.4% of their original triangle counts while preserving appearance, including foliage, animated characters, normal-map details, and cross-renderer conversions.
Key insights
- Visual fidelity can improve when geometric detail is moved into normal, displacement, or material textures; jointly optimizing these components lets the system determine how to divide information between geometry and appearance.
- Optimizing against rendered images preserves perceptual properties that geometry-only metrics miss, such as foliage density, transparency, specular behavior, and the appearance of animated assets across multiple poses.
- Initialization, topology, and texture parameterization remain important constraints: the current system does not optimize topology or UVs, and large geometric mismatches can produce view-dependent artifacts or slow convergence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.