Research map / 3D vision / Face and shape
Differentiable Micro-Mesh Construction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an end-to-end differentiable framework that builds micro-meshes (µ-meshes) from standard meshes by jointly optimizing base mesh offsets and displacements under rendering, Laplacian reparameterization, and shell-volume and isotropy regularizers.
Problem
Converting standard high-resolution meshes into micro-mesh (µ-mesh) format is done with stage-based pipelines (decimation, remeshing, displacement baking) that accumulate error and cannot directly optimize rendering fidelity or shell volume.
Approach
Treat µ-mesh construction as a single differentiable optimization: parameterize base-vertex offsets and per-micro-vertex displacements, reparameterize them with cotangent Laplacian diffusion, pre-tessellate for differentiable rasterization, and progressively subdivide with curriculum training; add shell-volume variance and isotropy regularizers and a visual-guided adaptive LOD criterion.
Key results
- On the extreme subset (>2M faces), the method reaches 44.89 PSNR and 0.083 shell volume versus 42.61 PSNR and 0.089 for the staged method of Maggiordomo et al.
- On the large subset, PSNR improves from 42.42 (staged) to 44.18 with shell volume 0.082 vs 0.086.
- Ablation: removing progressive scheme and reparameterization drops extreme-subset PSNR to 43.65 (from 44.89) and 42.25 without reparameterization path.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add progressive subdivision and optimization (ablation) | Differentiable Micro-Mesh Construction · Extreme subset | PSNR: 43.65 -> 44.89 | Helped |
| add Laplacian reparameterization (ablation) | Differentiable Micro-Mesh Construction · Extreme subset | PSNR: 42.25 -> 44.89 | Helped |
| add shell volume regularizer Lsv (ablation) | Differentiable Micro-Mesh Construction · Extreme subset | PSNR: 44.60 -> 44.89 | Helped |
| add isotropy regularizer Liso (ablation) | Differentiable Micro-Mesh Construction · Extreme subset | PSNR: 44.52 -> 44.89 | Helped |
| add adaptive level of detail (ablation) | Differentiable Micro-Mesh Construction · Extreme subset | PSNR: 44.78 -> 44.89 | Helped |
| add shell volume regularization Lsv (ablation) | Differentiable Micro-Mesh Construction · one mesh (Figure 8) | PSNR: 33.40 -> 33.65 | Helped |
Surprising findings
- Base mesh geometry error and isotropy only roughly correlate with final µ-mesh quality, so the authors focus on global µ-mesh objectives instead.
Limitations
- Construction time is comparable to or longer than the staged baseline on medium and large meshes (e.g., 3.9 min vs 2.0 min on medium).
- Base mesh quality metrics do not track final µ-mesh quality closely.
Experimental setup
Datasets: 121 high-resolution meshes inherited from Maggiordomo et al. (0.1M-4M faces)
Benchmarks: 121 high-resolution mesh models (medium, large, extreme subsets)
Metrics: rendering PSNR shell volume quadric geometry error isotropy construction time storage
Baselines: Simplygon Maggiordomo et al. staged micro-mesh pipeline
Builds on: micro-mesh Nicolet et al. diffusion reparameterization (Instant Field-Aware / Large Steps) nvdiffrast Loop subdivision Quadric Error Metric decimation
Closest papers by meaning
- LinPrim: Linear Primitives for Differentiable Volumetric Rendering (NeurIPS 2025 · similarity 0.84)
- Tetrahedron Splatting for 3D Generation (NeurIPS 2024 · similarity 0.84)
- UniGS: Unified Geometry-Aware Gaussian Splatting for Multimodal Rendering (NeurIPS 2026 · similarity 0.83)
- GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation (ICLR 2025 · similarity 0.83)
- DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation (NeurIPS 2024 · similarity 0.83)
- QUADify: Extracting Meshes with Pixel-level Details and Materials from Images (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.