Research map / 3D vision / 3D generation
VQ3D: Learning a 3D-Aware Generative Model on ImageNet
Kyle Sargent, Jing Yu Koh, Han Zhang, Huiwen Chang, Charles Herrmann, Pratul Srinivasan, Jiajun Wu, Deqing Sun
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 GAN-based 3D generator with vector-quantized autoencoder plus NeRF decoder | 3D-aware generative model · ImageNet generation | FID: 69.8 -> 16.8 | Improved |
| add pseudo-GT depth loss to baseline GAN methods (ablation) | pi-GAN, GIRAFFE, EG3D · ImageNet generation | FID: EG3D 109.3 -> 91.8; GIRAFFE 132.1 -> 132.0 | No clear effect |
| add weighted pointwise depth loss (ablation) | VQ3D Stage 2 · ImageNet generation | Depth accuracy: 1.90 -> 0.16 | Improved |
| remove depth loss from Stage 1 (ablation) | VQ3D Stage 1 · ImageNet reconstruction | FID / depth accuracy: FID 11.2 -> 4.0; depth acc. 0.18 -> 1.91 | Mixed |
| use triplanes instead of a fully implicit representation (ablation) | VQ3D Stage 1 · ImageNet reconstruction | FID: 273.5 -> 11.2 | Improved |
| use ViT encoder and decoder instead of CNN (ablation) | VQ3D Stage 1 · ImageNet reconstruction | training stability | Improved |
| add GAN loss and depth scale loss (ablation) | VQ3D Stage 1 · ImageNet reconstruction | disparity scale / novel view quality: disparity scale 1.27 -> 1.00 (GAN); 1.21 -> 1.00 (scale) | Improved |
| add NeRF distortion and interlevel losses (ablation) | VQ3D Stage 1 · ImageNet reconstruction | depth accuracy: 0.28 -> 0.18 | Improved |
| restrict sampling to top-k 1000 instead of the full 8192 codebook (ablation) | VQ3D Stage 2 · ImageNet generation | FID: 36.1 -> 31.5 | Improved |
| use overly restrictive top-p (0.95) (ablation) | VQ3D Stage 2 · ImageNet generation | FID: 32.2 (p=0.98) -> 35.7 (p=0.95) | Hurt |
| VQ3D vs best baselines on CompCars single-class dataset | VQ3D · CompCars | FID: 7.2 -> 7.3 (vs GIRAFFE HD) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- StyleNeRF — StyleNeRF: A Style-based 3D Aware Generator for High-resolution Image Synthesis (ICLR 2022) · better on 2 of 2 shared results
- pi-GAN — Pi-GAN: Periodic Implicit Generative Adversarial Networks for 3D-Aware Image Synthesis (CVPR 2021) · better on 2 of 2 shared results
- GIRAFFE — GIRAFFE: Representing Scenes As Compositional Generative Neural Feature Fields (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VFusion3D: Learning Scalable 3D Generative Models from Video Diffusion Models (ECCV 2024 · similarity 0.89)
- Generating Diverse High-Fidelity Images with VQ-VAE-2 (NeurIPS 2019 · similarity 0.88)
- SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE (CVPR 2025 · similarity 0.88)
- Can3Tok: Canonical 3D Tokenization and Latent Modeling of Scene-Level 3D Gaussians (ICCV 2025 · similarity 0.88)
- G3DR: Generative 3D Reconstruction in ImageNet (CVPR 2024 · similarity 0.87)
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (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.
