Research map / 3D vision / Face and shape
PGC: Physics-Based Gaussian Cloth from a Single Pose
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.
Reconstructs simulation-ready garments from a single static multi-view capture by combining mesh-embedded 3D Gaussian splats with a cloth-specific physically based shading model via high/low-frequency hybrid rendering.
Problem
Reconstructing photorealistic, simulation-ready garments with fine appearance that generalize to new poses and lighting without multi-frame tracking or large data.
Approach
Hybrid rendering that takes the low-pass of a PBR-shaded mesh (far-field pose-dependent shading, with sheen BRDF) and adds the high-pass of mesh-embedded Gaussian splats (near-field fabric detail), with Gaussian colors driven by the shaded texture, all fitted from one static frame.
Key results
- Full method achieves FSIM 0.834 and LPIPS 3.38e-2 on novel poses, beating Animatable Gaussians (0.827, 3.39) and SCARF (0.764, 5.00).
- Ablation: 3DGS-Only reaches FSIM 0.825 / LPIPS 3.41 and PBR-Only 0.809 / 4.67, so combining both representations gives the best scores.
- Method runs in real time for rendering and simulation, and avoids the 24.5 hours of multi-view registration needed by Gaussian Garments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PBR shading to 3DGS-only rendering (ablation) | PGC · Novel poses (4 video sequences) | FSIM: 0.825 -> 0.834 | Helped |
| Add PBR shading to 3DGS-only rendering (ablation) | PGC · Novel poses (4 video sequences) | LPIPS (×10−2): 3.41 -> 3.38 | Helped |
| Add Gaussian details to PBR-only rendering (ablation) | PGC · Novel poses (4 video sequences) | FSIM: 0.809 -> 0.834 | Helped |
| Add Gaussian details to PBR-only rendering (ablation) | PGC · Novel poses (4 video sequences) | LPIPS (×10−2): 4.67 -> 3.38 | Helped |
Surprising findings
- Animatable Gaussians gives slightly higher PSNR than the proposed method, attributed to sag in the simulated rest shape rather than appearance quality.
Limitations
- Ignores pose- and lighting-dependent effects on high-frequency appearance; limited by 3DGS coverage of under-observed areas.
- Albedo extraction per view yields inconsistent colors when back-projected, causing far-field artifacts.
- Rest shape taken from reconstructed geometry, causing sagging under gravity; shell geometry assumption fails for pockets, multi-layer garments, and openings.
Experimental setup
Datasets: in-house multi-view capture of 4 garments (170 cameras, 4096x2668)
Benchmarks: in-house novel-pose multi-view sequences (4 sequences, 150 unseen frames each)
Metrics: LPIPS FSIM
Baselines: SCARF Animatable Gaussians 3DGS-Only PBR-Only
Builds on: 3D Gaussian splatting Disney BRDF XPBD Gaussian Garments
Closest papers by meaning
- GBC-Splat: Generalizable Gaussian-Based Clothed Human Digitalization under Sparse RGB Cameras (CVPR 2025 · similarity 0.91)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- Gaussian Head & Shoulders: High Fidelity Neural Upper Body Avatars with Anchor Gaussian Guided Texture Warping (ICLR 2025 · similarity 0.87)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.87)
- GarFast: Realistic and Fast Garment Transfer with a Simplified Parser-Free Approach (AAAI 2025 · similarity 0.86)
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering (CVPR 2024 · similarity 0.86)
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.