Research map / 3D vision / Gaussian splatting
FlowR: Flowing from Sparse to Dense 3D Reconstructions
Tobias Fischer, Samuel Rota Bulò, Yung-Hsu Yang, Nikhil Keetha, Lorenzo Porzi, Norman Müller, Katja Schwarz, Jonathon Luiten, Marc Pollefeys, Peter Kontschieder
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 rendering-conditioned source distribution p0(z|y) instead of Gaussian noise (ablation) | FlowR flow matching · DL3DV140 | SSIM: 0.769 -> 0.808 | Improved |
| use rendering-conditioned source distribution instead of Gaussian noise (ablation) | FlowR flow matching · ScanNet++ | PSNR: 24.09 -> 25.53 | Improved |
| use rendering-conditioned source distribution instead of Gaussian noise (ablation) | FlowR flow matching · DL3DV140 | LPIPS: 0.200 -> 0.180 | Improved |
| add alternative target-view loss L_tgt (L2, SSIM, LPIPS) instead of vanilla L_GS (ablation) | FlowR refined reconstruction · DL3DV140 | LPIPS: 0.279 -> 0.248 | Improved |
| add re-initialization of pointcloud with generated views (ablation) | FlowR refined reconstruction · DL3DV140 | PSNR: 23.72 -> 23.78 | Improved |
| add re-initialization with generated views in 24-view split (ablation) | FlowR refined reconstruction · DL3DV140 (24-view) | PSNR: 25.10 -> 25.13 | Improved |
| add MASt3R dense pointcloud initialization instead of Splatfacto default (ablation) | initial reconstruction · DL3DV140 | PSNR: 19.44 -> 21.72 | Improved |
| add co-visibility graph G_vis (ablation) | initial reconstruction · DL3DV140 | view pairs: 342 -> 129 | Improved |
| add re-triangulation of matches (ablation) | initial reconstruction · DL3DV140 | PSNR: 21.76 -> 22.23; LPIPS 0.293 -> 0.322 | Mixed |
| add adaptive density control with 5K steps (ablation) | initial reconstruction · DL3DV140 | LPIPS: 0.322 -> 0.292 | Improved |
| train longer (30K instead of 5K steps) (ablation) | initial reconstruction · DL3DV140 | PSNR: 22.58 -> 22.21 | No clear effect |
| add flow refinement of test views (FlowR++) on top of FlowR (ablation) | FlowR · ScanNet++ | LPIPS: 0.303 -> 0.250 | Improved |
Later papers that compared with this method
1 later accepted paper reports FlowR as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SyncFix: Multi-View Consistent Diffusion Refinement of 3D Reconstructions (ECCV 2026 · better on 7 of 7 shared results)
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.
- ViewCrafter — ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis (TPAMI 2025) · better on 4 of 4 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
- GenRec: Knowing Where to Reconstruct and Where to Generate (NeurIPS 2026 · similarity 0.90)
- SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 2025 · similarity 0.89)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion (NeurIPS 2025 · similarity 0.88)
- MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.88)
- Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations (CVPR 2025 · 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.
