Research map / 3D vision / 3D generation
Native Spatio-Temporal 4D Variational Autoencoder
Lihe Ding, Weicai Ye, Shaocong Dong, Xintao Wang, Pengfei Wan, Kun Gai, Tianfan Xue
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 |
|---|---|---|---|
| add native spatio-temporal window attention instead of decoupled 3D+1D convolutions (ablation) | Native 4D VAE · 4D-Dyn | CD-L1: 0.0025 -> 0.0003 | Improved |
| add native spatio-temporal window attention (ablation) | Native 4D VAE · 4D-Dyn | PSNR: 27.63 -> 29.63 | Improved |
| add differentiable voxel rendering loss (ablation) | Native 4D VAE · 4D-Dyn | PSNR: 25.08 -> 29.63 | Improved |
| add differentiable voxel rendering loss (ablation) | Native 4D VAE · 4D-Dyn | CD-L1: 0.0011 -> 0.0003 | Improved |
| add Plücker camera embedding in encoder (ablation) | Native 4D VAE · 4D-Dyn | PSNR: 29.07 -> 29.63 | Improved |
| add Plücker camera embedding in encoder (ablation) | Native 4D VAE · 4D-Dyn | CD-L1: 0.0004 -> 0.0003 | Improved |
| add mesh pre-processing before voxelization (ablation) | Native 4D VAE · 4D-Dyn | Precision@0.0005: 0.8579 -> 0.9876 | Improved |
| strengthen PointMapVAE with VGGT transformer backbone (ablation) | Native 4D VAE · 4D-Dyn | CD-L1: 0.019 -> 0.008 (still behind native 4D VAE at 0.0003) | Improved |
| use native 4D voxel VAE instead of flow-based DyMeshVAE | Native 4D VAE · DyMesh dataset | Chamfer-Distance-L1: 0.0008 -> 0.0001 | Improved |
| use native 4D voxel VAE instead of PointMapVAE | Native 4D VAE · 4D-Dyn | CD-L1: 0.024 -> 0.0003 | Improved |
Closest papers by meaning
- 4DWeaver: Bridging Reconstruction and Generation Via Compact Autoregressive Priors (TPAMI 2026 · similarity 0.88)
- Diffusion4D: Fast Spatial-temporal Consistent 4D generation via Video Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Fused View-Time Attention and Feedforward Reconstruction for 4D Scene Generation (NeurIPS 2025 · similarity 0.87)
- Improved Video VAE for Latent Video Diffusion Model (CVPR 2025 · similarity 0.87)
- SV4D: Dynamic 3D Content Generation with Multi-Frame and Multi-View Consistency (ICLR 2025 · similarity 0.86)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · 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.
