Research map / Image generation / Image restoration
Era3D: High-Resolution Multiview Diffusion using Efficient Row-wise Attention
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.
Era3D generates 512x512 canonical orthogonal multiview normal and color images from a single arbitrary-camera input, using a regression-and-condition module for focal length and elevation and efficient row-wise multiview attention.
Problem
Single-view 3D reconstruction via multiview diffusion suffers from perspective distortion when inputs use camera intrinsics different from the assumed fixed camera, and dense multiview attention becomes prohibitively costly at high resolution.
Approach
Generate multiview images in a canonical orthogonal camera at 0° elevation while allowing arbitrary-intrinsic inputs; a UNet-feature regression head (EFReg) predicts input elevation and focal length and conditions the denoiser; row-wise attention exploits the fact that epipolar lines become image rows in this canonical setup.
Key results
- On GSO, Era3D achieves CD 0.0217 and IoU 0.5973 versus Wonder3D's 0.0248 and 0.5678.
- Row-wise attention at 512 resolution uses 1.66 GB and 2.23 ms versus 35.32 GB and 220.41 ms for dense multiview attention without xFormers.
- Elevation prediction error of 2.69 degrees versus 10.24 for DINOv2 baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EFReg camera regression and condition on predicted elevation and focal length during denoising (ablation) | Era3D · GSO single-view reconstruction | 3D reconstruction quality | Helped |
| Replace dense multiview attention with row-wise attention (ablation) | Era3D · GSO, 256 resolution | CD: 0.0239 -> 0.0232 | Helped |
| Replace dense multiview attention with row-wise attention (ablation) | Era3D · GSO, 256 resolution | IoU: 0.5877 -> 0.5831 | Hurt |
| Replace dense multiview attention with row-wise attention (ablation) | Era3D · GSO, 256 resolution | LPIPS: 0.140 -> 0.137 | Helped |
| Replace dense multiview attention with row-wise attention (ablation) | Era3D · GSO, 256 resolution | PSNR: 20.73 -> 20.92 | Helped |
| Replace dense multiview attention with row-wise attention (ablation) | Era3D · GSO, 256 resolution | SSIM: 0.819 -> 0.813 | Hurt |
| Replace dense attention with row-wise attention (ablation) | Era3D · 256 resolution, without xFormers | memory usage (G): 3.02 -> 0.95 | Helped |
| Replace dense attention with row-wise attention (ablation) | Era3D · 512 resolution, without xFormers | memory usage (G): 35.32 -> 1.66 | Helped |
| Replace dense attention with row-wise attention (ablation) | Era3D · 512 resolution, without xFormers | running time (ms): 220.41 -> 2.23 | Helped |
| Integrate cross-domain fusion into the row-wise multiview attention block rather than the self-attention block (ablation) | Era3D · cross-domain image generation | normal-map detail | Hurt |
| Use a linear noise scheduler instead of the scaled linear scheduler (ablation) | Era3D · 512 resolution generation | generation quality | Helped |
Surprising findings
- Row-wise attention at 256 resolution slightly outperforms dense multiview attention on Chamfer distance, LPIPS and PSNR despite fewer attention tokens.
Limitations
- Generates only 6 multiview images, which struggles with thin or intricate geometry.
- Neural SDF reconstruction cannot produce open surfaces.
- Focal length prediction is limited for large focal lengths and the authors note room for improvement.
Experimental setup
Datasets: Objaverse Google Scanned Objects
Benchmarks: Google Scanned Objects
Metrics: LPIPS SSIM PSNR Chamfer Distance Volume IoU elevation error normalized focal length error
Baselines: RealFusion Zero-1-to-3 One-2-3-45 Shap-E Magic123 SyncDreamer Wonder3D LGM Unique3D
Builds on: Stable Diffusion Wonder3D Zero-1-to-3 epipolar attention NeuS
Closest papers by meaning
- EpiDiff: Enhancing Multi-View Synthesis via Localized Epipolar-Constrained Diffusion (CVPR 2024 · similarity 0.90)
- MEAT: Multiview Diffusion Model for Human Generation on Megapixels with Mesh Attention (CVPR 2025 · similarity 0.90)
- Efficient-3Dim: Learning a Generalizable Single-image Novel-view Synthesizer in One Day (ICLR 2024 · similarity 0.89)
- 3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement (CVPR 2025 · similarity 0.88)
- ViVid-1-to-3: Novel View Synthesis with Video Diffusion Models (CVPR 2024 · similarity 0.88)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · 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.