Research map / 3D vision / 3D generation
UpFusion: Novel View Diffusion from Unposed Sparse View Observations
Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, Shubham Tulsiani
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 |
|---|---|---|---|
| use pose-free UpFusion instead of SparseFusion with RelPose++ estimated poses | UpFusion (3D) · Co3Dv2 seen, 6 views | PSNR-A: 17.12 -> 18.96 | Improved |
| add more unposed input views (1 to 6) | UpFusion (3D) · Co3Dv2 seen | PSNR-A: 18.17 -> 18.96 | Improved |
| add more views to SparseFusion with RelPose++ poses | SparseFusion (RelPose++) · Co3Dv2 seen | PSNR-A: 17.76 -> 17.12 (3V to 6V) | Hurt |
| distill a 3D-consistent NeRF via score distillation from the 2D diffusion model | UpFusion (2D) · Co3Dv2 seen, 3 views | PSNR-A: 17.12 -> 18.68 | Improved |
| use diffusion (UpFusion 2D) instead of regression (UpSRT) | UpSRT · Co3Dv2 seen, 3 views | LPIPS-A: 0.32 -> 0.22 | Improved |
| use diffusion (UpFusion 2D) instead of regression (UpSRT) | UpSRT · Co3Dv2 seen, 3 views | PSNR-A: 17.75 -> 17.12 | Hurt |
| add set-latent attention conditioning to decoder features (ablation) | UpFusion 2D (teddybear) · Co3Dv2 teddybear, 3 views | PSNR-A: 15.28 -> 16.11 | Improved |
| add decoder features to set-latent conditioning (ablation) | UpFusion 2D (teddybear) · Co3Dv2 teddybear, 3 views | PSNR-A: 13.38 -> 16.11 | Improved |
| evaluate on unseen categories | UpFusion (3D) · Co3Dv2 unseen vs seen, 3 views | PSNR-A: 18.68 (seen) vs 18.83 (unseen) | No clear effect |
| use six unposed views instead of pose-optimizing FORGE | UpFusion-dagger (3D) · GSO, 6 views | PSNR: 17.40 -> 22.51 | Improved |
| use UpFusion-dagger with a single view instead of Zero-1-to-3 (threestudio) | single-view 3D · GSO, 1 view | PSNR: 21.71 -> 20.52 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- One-2-3-45 — One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization (NeurIPS 2023) · 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
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.91)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.90)
- ViewFusion: Towards Multi-View Consistency via Interpolated Denoising (CVPR 2024 · similarity 0.89)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.89)
- Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion (NeurIPS 2025 · similarity 0.89)
- LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.89)
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.
