Research map / Point clouds / Shape and surface completion
Class-Partitioned VQ-VAE and Latent Flow Matching for Point Cloud Scene Generation
Dasith de Silva Edirimuni, Ajmal Saeed Mian
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 Diffuscene's diffusion with LFMM latent flow matching, keeping a standard VAE decoder | Diffuscene · 3D-FRONT living room generation | Chamfer distance x10^3: 30.63 -> 24.65 | Improved |
| add CPVQ-VAE decoder to LFMM instead of standard VAE (ablation) | LFMM · 3D-FRONT living room generation | Chamfer distance x10^3: 24.65 -> 9.06 | Improved |
| add CPVQ-VAE decoder to LFMM instead of standard VAE (ablation) | LFMM · 3D-FRONT bedroom generation | Chamfer distance x10^3: 4.24 -> 2.46 | Improved |
| add class partitioning to a plain VQ-VAE (ablation) | VQ-VAE · 3D-FRONT bedroom autoencoder ablation | Chamfer distance x10^3: 36.27 -> 5.00 | Improved |
| add class-aware running average update to class-partitioned VQ-VAE (ablation) | CPVQ-VAE · 3D-FRONT bedroom autoencoder ablation | Chamfer distance x10^3: 5.00 -> 2.46 | Improved |
| use plain VQ-VAE instead of standard VAE (ablation) | autoencoder · 3D-FRONT bedroom | Chamfer distance x10^3: 4.24 -> 36.27 | Hurt |
| increase sampling steps from 10 to 100 (ablation) | CPVQ-VAE+LFMM · 3D-FRONT bedroom retrieval | KID x10^3: 1.324 -> 0.794 | Improved |
| increase sampling steps from 100 to 1000 (ablation) | CPVQ-VAE+LFMM · 3D-FRONT bedroom retrieval | Runtime (s): 0.892 -> 8.963 | Hurt |
| increase sampling steps from 100 to 1000 (ablation) | CPVQ-VAE+LFMM · 3D-FRONT bedroom retrieval | FID: 21.34 -> 21.75 | No clear effect |
| use LFMM flow matching instead of Diffuscene diffusion | scene generation · 3D-FRONT all scene types | Runtime (s): 9.153 -> 0.892 | Improved |
| use CPVQ-VAE instead of standard VAE for retrieval (ablation) | LFMM · 3D-FRONT living room retrieval | FID: 24.46 -> 26.10 | Hurt |
Closest papers by meaning
- Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022 · similarity 0.88)
- Representing 3D Shapes with 64 Latent Vectors for 3D Diffusion Models (ICCV 2025 · similarity 0.88)
- Hierarchy Denoising Recursive Autoencoders for 3D Scene Layout Prediction (CVPR 2019 · similarity 0.87)
- TopNet: Structural Point Cloud Decoder (CVPR 2019 · similarity 0.87)
- Can3Tok: Canonical 3D Tokenization and Latent Modeling of Scene-Level 3D Gaussians (ICCV 2025 · similarity 0.87)
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 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.
