Research map / 3D vision / Depth and stereo
RayZer: A Self-supervised Large View Synthesis Model
Hanwen Jiang, Hao Tan, Peng Wang, Haian Jin, Yue Zhao, Sai Bi, Kai Zhang, Fujun Luan, Kalyan Sunkavalli, Qixing Huang, Georgios Pavlakos
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 Plucker ray map conditioning of the scene reconstructor (ablation) | RayZer · DL3DV (even sample) | PSNR: 22.73 -> 24.36 | Improved |
| use explicit SE(3) poses as an information bottleneck instead of latent camera tokens (ablation) | RayZer · DL3DV (even sample) | PSNR: 23.13 -> 22.73 without Plucker rays (latent tokens slightly better but leak target info and cannot be interpolated) | Mixed |
| predict pose first then scene instead of scene first (ablation) | RayZer · DL3DV (even sample) | PSNR: 13.31 -> 24.36 | Improved |
| use latent set scene representation instead of 3DGS with rasterization (ablation) | RayZer · DL3DV | PSNR: 3DGS variant failed to converge | Improved |
| select middle frame rather than first frame as canonical view (ablation) | RayZer · DL3DV (even sample) | PSNR: 23.86 -> 24.36 | Improved |
| add frame-sampling curriculum from small to large baselines (ablation) | RayZer · DL3DV (even sample) | PSNR: 23.87 -> 24.36 | Improved |
| train on continuous video frames with image index embeddings instead of shuffled unordered images (ablation) | RayZer · DL3DV (even sample) | PSNR: 20.56 -> 24.36 | Improved |
| self-supervised RayZer versus supervised LVSM trained on COLMAP poses | RayZer · DL3DV (even sample) | PSNR: 23.69 -> 24.36 | Improved |
| self-supervised RayZer versus supervised LVSM trained on COLMAP poses | RayZer · RealEstate (even sample) | PSNR: 27.00 -> 27.48 | Improved |
| self-supervised RayZer versus oracle LVSM with perfect Blender poses | RayZer · Objaverse (even sample) | PSNR: 32.34 -> 31.52 | Hurt |
| self-supervised pre-trained pose encoder with MLP probe versus supervised from scratch | RayZer · DL3DV | R@10 rotation accuracy (%): 39.3 -> 47.6 | Improved |
| self-supervised pre-trained pose encoder with MLP probe versus supervised from scratch | RayZer · Objaverse | R@10 rotation accuracy (%): 19.8 -> 33.6 | Improved |
Later papers that compared with this method
7 later accepted papers report RayZer as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- True Self-Supervised Novel View Synthesis is Transferable (ICLR 2026 · better on 8 of 8 shared results)
- REON-NVS: Real-Time Online Novel-View Synthesis from Sparse-View Videos (ECCV 2026 · better on 7 of 7 shared results)
- StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views (ECCV 2026 · better on 3 of 3 shared results)
- RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video (ECCV 2026 · better on 2 of 2 shared results)
- From Rays to Projections: Better Inputs for Feed-Forward View Synthesis (CVPR 2026 · better on 1 of 1 shared result)
- Reliev3R: Relieving Feed-forward 3D Reconstruction from Multi-View Geometric Annotations (CVPR 2026 · better on 1 of 1 shared result)
- SceneTok: A Compressed, Diffusable Token Space for 3D Scenes (CVPR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- GS-LRM — GS-LRM: Large Reconstruction Model for 3D Gaussian Splatting (ECCV 2024) · better on 8 of 8 shared results
- PF-LRM — PF-LRM: Pose-Free Large Reconstruction Model for Joint Pose and Shape Prediction (ICLR 2024) · better on 3 of 3 shared results
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
- RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video (ECCV 2026 · similarity 0.92)
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 2026 · similarity 0.87)
- Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion (CVPR 2025 · similarity 0.86)
- LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.86)
- RaySt3R: Predicting Novel Depth Maps for Zero-Shot Object Completion (NeurIPS 2025 · similarity 0.86)
- DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features (NeurIPS 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.
