Research map / Image generation / Fast diffusion sampling
Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation
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.
Applies diffusion, flow matching, and Riemannian flow matching on the sphere to global visual geolocation, enabling state-of-the-art accuracy and probabilistic location densities with new metrics for probabilistic geolocation.
Problem
Global visual geolocation predicts where an image was taken; existing methods are deterministic and ignore the varying localizability (ambiguity) of images.
Approach
Treat geolocation as a generative denoising process over coordinates conditioned on image embeddings, using Riemannian flow matching so denoising stays on the Earth's sphere, and compute exact log-likelihoods of any location via the instantaneous change-of-variables formula.
Key results
- RFM S2 reaches 3767 GeoScore and 76.2% country accuracy on OSV-5M, beating the hybrid approach of Astruc et al. (GeoScore +406, distance -745 km).
- On YFCC4k, RFM10M S2 achieves 33.5% accuracy at 25 km and 77.7% at 2500 km, outperforming PIGEON (24.4% at 25 km).
- On probabilistic geolocation, RFlowMatch S2 achieves NLL of -1.51 on OSV-5M and the best precision/density/coverage among compared models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Riemannian flow matching on the sphere instead of the Hybrid baseline | Visual geolocation · OSV-5M test | GeoScore: 3361 -> 3767 | Helped |
| Use Riemannian flow matching on the sphere instead of SC Retrieval | Visual geolocation · iNat21 validation | country accuracy (%): 45.8 -> 44.2 | Hurt |
| Train RFM S2 for 9 million iterations instead of one million (ablation) | RFM S2 · YFCC-4k | GeoScore: 2889 -> 3210 | Helped |
| Use flow matching in R3 instead of diffusion in R3 | Generative visual geolocation · OSV-5M test | GeoScore: 3762 -> 3688 | Hurt |
| Constrain flow matching to the sphere with Riemannian flow matching instead of flow matching in R3 | Generative visual geolocation · OSV-5M test | GeoScore: 3688 -> 3767 | Helped |
| Use Riemannian flow matching on S2 instead of flow matching in R3 | Probabilistic visual geolocation · iNat21 | precision: 0.76 -> 0.88 | Helped |
| Replace a single vMF distribution with a mixture of vMF distributions | vMF baseline · iNat21 | NLL: 1.99 -> -0.23 | Helped |
| Increase RFM inference from 1 timestep to 16 timesteps (ablation) | RFM S2 · OSV-5M | GeoScore: 591 -> 3744 | Helped |
| Increase RFM inference beyond 16 timesteps (ablation) | RFM S2 · OSV-5M | GeoScore: 3744; 3746 | No clear effect |
| Increase classifier-free guidance scale (ablation) | Generative visual geolocation · OSV-5M | GeoScore | Helped |
| Increase classifier-free guidance scale (ablation) | Probabilistic visual geolocation · OSV-5M | precision and recall | Hurt |
Surprising findings
- Flow matching consistently outperforms diffusion for geolocation.
- A single-component vMF regression model performs similarly to a discriminative regression baseline, while vMF mixtures overfit the training set.
Limitations
- Retrieval-based methods retain an advantage at extremely fine-grained (sub-25 km) resolutions.
- Guidance scale trades off geolocation accuracy against coverage of the true distribution.
- Likelihoods of R3 and S2 models are not directly comparable.
Experimental setup
Datasets: OpenStreetView-5M iNat21 YFCC-100M YFCC4k
Benchmarks: OpenStreetView-5M iNat21 YFCC4k
Metrics: Haversine distance GeoScore accuracy at country/region/city/distance thresholds negative log-likelihood precision recall density coverage localizability
Baselines: StreetCLIP zero-shot Regression (Astruc et al.) ISNs Hybrid (Astruc et al.) SC Retrieval PlaNet CPlaNet Translocator GeoDecoder PIGEON vMF vMFMix
Builds on: denoising diffusion implicit models flow matching Riemannian flow matching classifier-free guidance PIGEON
Closest papers by meaning
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.89)
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.89)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.88)
- LocDiff: Identifying Locations on Earth by Diffusing in the Hilbert Space (NeurIPS 2025 · similarity 0.88)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.87)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · 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.