Research map / 3D vision / Face and shape
DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation
Eungjune Shim, Hansol Lee, Eunjung Ju
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace binary occupancy with continuous TSDF (no normal loss) (ablation) | DiffGI-VAE · GarmageSet | CD (x1e-3): 1.503 -> 0.595 | Improved |
| replace binary occupancy with continuous TSDF (no normal loss) (ablation) | DiffGI-VAE · GarmageSet | JSD (x1e-3): 4.539 -> 2.169 | Improved |
| add geometry-aware normal rendering loss (ablation) | DiffGI-VAE (TSDF) · GarmageSet | NC: 0.921 -> 0.961 | Improved |
| add geometry-aware normal rendering loss (ablation) | DiffGI-VAE (TSDF) · GarmageSet | CD (x1e-3): 0.595 -> 0.461 | Improved |
| add normal rendering loss to occupancy representation (ablation) | DiffGI-VAE (Occ) · GarmageSet | NC: 0.906 -> 0.947 | Improved |
| initialize VAE from SD1.5 instead of random (ablation) | DiffGI-VAE · GarmageSet | CD (x1e-3): 0.47 -> 0.46 | No clear effect |
| compress 256x256 geometry image into 32x32x4 latent vs uncompressed 64x64x4 Omages | DiffGI-VAE · ABO | NC: 0.89 -> 0.83 | Hurt |
| compress into 32x32x4 latent vs Omages | DiffGI-VAE · ABO | CD (x1e-3): 0.89 -> 0.83 | Improved |
| use DiffGI diffusion vs Omages for label-conditioned generation | DiffGI · ABO | mean P-FID: 31.14 -> 19.91 | Improved |
| use DiffGI diffusion vs Omages for Table category | DiffGI · ABO Table category | P-KID: 0.10 -> 0.11 | No clear effect |
| geometric UV augmentation (3,800 -> 500K samples) for larger DiT (ablation) | DiT-Base · ABO | training stability | Improved |
| TSDF vs occupancy at resolution 64/128 without VAE (ablation) | geometry image · ABO and GarmageSet | Hausdorff distance | Improved |
Closest papers by meaning
- NI-Tex: Non-isometric Image-based Garment Texture Generation (CVPR 2026 · similarity 0.87)
- Ghost on the Shell: An Expressive Representation of General 3D Shapes (ICLR 2024 · similarity 0.87)
- TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer (CVPR 2025 · similarity 0.86)
- Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields (CVPR 2025 · similarity 0.86)
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.85)
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.
