Research map / Image generation / Autoregressive and latent generation
Representing 3D Shapes with 64 Latent Vectors for 3D Diffusion Models
In Cho, Youngbeom Yoo, Subin Jeon, Seon Joo Kim
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 |
|---|---|---|---|
| use patch-based progressive encoder instead of VecSet cross-attention encoder (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 91.2 -> 96.1 | Improved |
| add more cross-attention layers to the VecSet encoder (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 93.4 -> 93.5 | No clear effect |
| use triplane decoder instead of cross-attention decoder (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | Recon. throughput (sample/s): 3.43 -> 42.68 | Improved |
| use 3 cross-attention layers in decoder (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 96.1 -> 96.3 | Mixed |
| increase uncertainty-guided token pruning ratio to 90% (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 96.2 -> 95.9 | No clear effect |
| increase pruning ratio to 75% (R=128) (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | Training speed (iter/s): 4.21 -> 5.60 | Improved |
| add two-stage training (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 94.4 -> 96.1 | Improved |
| use reconstruction-based uncertainty instead of confidence-based (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 95.3 -> 96.1 | Improved |
| use learnable latent positions instead of input-dependent (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 96.0 -> 96.1 | No clear effect |
| increase number of latent vectors from 32 to 128 (ablation) | COD-VAE · ShapeNet reconstruction ablation (autoencoder M=32) | IoU (%): 96.1 -> 96.7 | Improved |
| COD-VAE with 64 latents versus VecSet with 64 latents | VecSet · ShapeNet reconstruction (autoencoder) | IoU (%): 91.2 -> 96.5 | Improved |
| COD-VAE M=64 versus VecSet M=1024 on Objaverse | VecSet · Objaverse reconstruction (VAE) | IoU (%): 79.8 -> 79.9 | Improved |
Closest papers by meaning
- LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion (ICLR 2025 · similarity 0.89)
- Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling (NeurIPS 2025 · similarity 0.89)
- Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders (CVPR 2025 · similarity 0.88)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models (ICLR 2025 · similarity 0.88)
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.
