Research map / 3D vision / NeRF and radiance fields
G-NeRF: Geometry-enhanced Novel View Synthesis from Single-View Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes G-NeRF, which distills geometry priors from a pretrained EG3D via truncated-latent multi-view synthesis and trains a NeRF-based single-view novel view synthesizer with a depth-aware discriminator.
Problem
Single-shot novel view synthesis from a single-view image without multi-view training data or test-time fine-tuning, where geometry priors are hard to learn without multi-view supervision.
Approach
Uses an off-the-shelf 3D GAN (EG3D) with latent truncation (psi=0.5) to synthesize multi-view triplets with depth as a free geometry prior, then trains a NeRF generator on mixed synthetic and real single-view images with a depth-aware discriminator that supervises depth maps against the GAN's depth.
Key results
- On FFHQ at 64^2, G-NeRF reaches FID 26.04 vs. 32.44 for Pix2NeRF, and on AFHQv2-Cats FID 18.64 vs. 25.34.
- On ShapeNet Cars & Chairs at 128^2, G-NeRF averages SSIM 0.87 and PSNR 19.87 vs. Pix2NeRF 0.81 / 16.99.
- Inference takes 1.1s versus 76.9s for PTI fine-tuning-based GAN inversion, with comparable FID (40.24 vs 35.53) and better depth (0.35 vs 0.37).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add G-NeRF geometry-guided synthesis and depth-aware training | Pix2NeRF on FFHQ · FFHQ 642 | FID: 32.44 -> 26.04 | Helped |
| add G-NeRF geometry-guided synthesis and depth-aware training | Pix2NeRF on CelebAMask-HQ · CelebAMask-HQ 642 | FID: 89.79 -> 75.76 | Helped |
| add G-NeRF geometry-guided synthesis and depth-aware training | Pix2NeRF on AFHQv2-Cats · AFHQv2-Cats 642 | FID: 25.34 -> 18.64 | Helped |
| add G-NeRF geometry-guided synthesis and depth-aware training | Pix2NeRF on ShapeNet Chairs · ShapeNet Chairs | SSIM: 0.80 -> 0.88 | Helped |
| add G-NeRF geometry-guided synthesis and depth-aware training | Pix2NeRF on ShapeNet Cars · ShapeNet Cars | SSIM: 0.73 -> 0.86 | Helped |
| add synthetic multi-view data (ablation) | G-NeRF on FFHQ · FFHQ | Depth: 0.42 -> 0.35 | Helped |
| use truncation ratio 0.5 instead of 1.0 for synthetic data (ablation) | G-NeRF on FFHQ · FFHQ | Depth: 0.37 -> 0.35 | Helped |
| add the depth-aware discriminator (ablation) | G-NeRF on FFHQ side faces · FFHQ side faces | Depth†: 0.59 -> 0.53 | Helped |
| add the depth-aware discriminator (ablation) | G-NeRF on FFHQ · FFHQ | SSIM: 0.64 -> 0.63 | Hurt |
| add the depth-aware discriminator (ablation) | G-NeRF on FFHQ · FFHQ | Depth: 0.35 -> 0.35 | No clear effect |
Surprising findings
- Without truncation (psi=1.0) the depth accuracy drops sharply and the model yields unrealistic geometry, showing that diversity-maximizing synthetic data harms geometry priors.
Limitations
- Relies on a pretrained 3D GAN (EG3D) as a source of geometry priors, so quality is bounded by that teacher.
- Fails to produce accurate geometry when one ear is not visible in the input without the depth-aware discriminator; depth evaluation is only on pseudo-ground-truth depth.
- 512 resolution results rely on a super-resolution module, so depth is not evaluated at that resolution.
Experimental setup
Datasets: FFHQ AFHQv2-Cats CelebAMask-HQ ShapeNet Cars ShapeNet Chairs
Benchmarks: FFHQ CelebAMask-HQ AFHQv2-Cats ShapeNet Cars ShapeNet Chairs
Metrics: FID KID Depth MSE ID (ArcFace cosine similarity) SSIM PSNR LPIPS
Baselines: Pix2NeRF ENR SRN PixelNeRF CodeNeRF VisionNeRF PTI MiDaS-based depth supervision Conditional EG3D
Builds on: EG3D Pix2NeRF pivotal tuning inversion StyleGAN
Closest papers by meaning
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.88)
- Generalizable Novel-View Synthesis using a Stereo Camera (CVPR 2024 · similarity 0.87)
- Unifying Correspondence Pose and NeRF for Generalized Pose-Free Novel View Synthesis (CVPR 2024 · similarity 0.87)
- Global and Hierarchical Geometry Consistency Priors for Few-shot NeRFs in Indoor Scenes (CVPR 2024 · similarity 0.87)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.87)
- NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis (NeurIPS 2026 · similarity 0.87)
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.