Image geolocalization predicts the location at which a photo was taken. Accurate geolocalization requires discriminating between both physically and semantically adjacent areas on Earth. However, prior diffusion-based approaches predict locations by generating coordinates, which are semantically unstable over space. We avoid this limitation by learning geolocalization behavior in the pretrained representation spaces of location encoders, which provide both positional and semantic understanding. To make this space usable for coordinate prediction, we introduce a Location Autoencoder (LAE), which pairs a frozen location encoder with a learned decoder mapping embeddings back to coordinates with kilometer-level accuracy worldwide. In the LAE space, we train DIG, a latent diffusion model that transports noise to location embeddings conditioned on image features. DIG consistently outperforms coordinate-based generative models and sets a new state of the art for parametric geolocalization on three benchmarks: OSV-5M, Im2GPS3k, and YFCC4k. Notably, DIG is the first generative model to outperform retrieval-based methods, and further improves retrieve-and-rerank systems when used as a prior. Code and models will be released.
Figures & tables
Figure 1: We propose D iffusion for I mage G eolocalization ( DIG ) in the location representation space. Conditioned on the image, a rectified flow transports noise to location embeddings, which a learned decoder maps to coordinates on the globe.
Figure 2
Figure 4: Overview of our approach. (a) Location Autoencoder (LAE) pretraining: a decoder learns to recover coordinates from the embeddings of a frozen location encoder. (b) A latent diffusion model (DIG), conditioned on the image, generates location embeddings, which the frozen LAE decoder maps back to coordinates. We leave out timestep conditioning for visual simplicity.
GeoS. ↑
Dist. ↓
Accuracy (%) ↑
Method
(km)
Ctry.
Reg.
City
Deterministic
SC 0-shot ( Haas et al., 2023 )
2273
2854
38.4
20.8
14.8
Regression ( Astruc et al., 2024 )
3028
1481
56.5
16.3
0.7
ISNs ( Muller-Budack et al., 2018 )
3331
2308
66.8
39.4
4.2
Hybrid ( Astruc et al., 2024 )
3361
1814
68.0
39.4
5.9
SC Retr. ( Haas et al., 2023 )
3597
1386
73.4
45.8
19.9
Table 1: Geolocation performance on OSV-5M ( Astruc et al., 2024 ) with official training and test splits. Dist. is mean distance. Bold indicates best and italics second best.
Im2GPS3k
YFCC4k
Method
1 km
25 km
200 km
750 km
2500 km
1 km
25 km
200 km
750 km
2500 km
Deterministic
PlaNet ( Weyand et al., 2016 )
8.5
24.8
34.3
48.4
64.6
5.6
14.3
22.2
36.4
55.8
CPlaNet ( Seo et al., 2018 )
10.2
26.5
34.6
48.6
64.6
7.9
14.8
21.9
36.4
55.5
Translocator ( Pramanick et al., 2022 )
11.8
31.1
46.7
58.9
80.1
8.4
18.6
27.0
41.1
60.4
GeoDecoder ( Clark et al., 2023 )
12.8
33.5
45.9
61.0
76.1
10.3
24.4
33.9
50.0
68.7
PIGEOTTO Haas et al. (2024)
11.3
36.7
53.8
72.4
85.3
10.4
23.7
40.6
62.2
77.7
Table 2: Geolocation performance on Im2GPS3k and YFCC4k. Accuracy (%) within distance thresholds, with best results in bold . We train different DIG-XL variants on MP16-Pro and YFCC-44M. 10M denotes training for 10M steps; ‡ denotes reimplementation with public weights.
Figure 5: Sampling in the LAE space. For each query (left) , we show three sampling trajectories as the cosine similarity between the latents and location embeddings across the globe (middle) , and display the top images retrieved with the generated embeddings, including their error in km (right) .
Table 3: DIG ablations. We train DIG-B models on OSV-5M using GeoCLIP LAEs; the highlighted row in each panel is our chosen configuration.
Figure 8
Decoder
< 0.01 km
< 0.1 km
< 0.5 km
GFLOPs
SatCLIP
0.10
9.31
89.0
0.076
GeoCLIP
0.01
1.23
26.1
0.076
TTE
0.09
8.70
85.6
0.076
+ 5 refinement steps
SatCLIP
90.2
99.4
100.0
0.161
GeoCLIP
99.7
99.8
99.9
0.737
Table 5: Sub-km reconstruction.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: ICA of location embeddings over land. Top: spherical harmonics with L=10,20,40 . Bottom: learned encoders SatCLIP, TTE, and GeoCLIP. Each panel maps its first three independent components to RGB.
Encoder
Input dim
Params
Steps
Locations seen
Density (/km 2 )
GeoCLIP
512
38.2M
6.1M
2.50×1010
167
TTE
512
38.2M
1.5M
6.14×109
41
SatCLIP
256
38.0M
400k
1.64×109
11
SH ( L=10 )
100
37.9M
200k
8.19×108
5.5
SH ( L=20 )
400
38.1M
200k
8.19×108
5.5
SH ( L=40 )
1600
39.0M
200k
8.19×108
5.5
Appendix
Table 6: Official decoder training. All decoders have depth 8 and use batch size 4096, with locations drawn uniformly over land. Density is locations seen per km 2 of land ( 1.50×108 km 2 ).
GeoCLIP
SatCLIP
TTE
Decoder
1 km
25 km
200 km
750 km
2500 km
1 km
25 km
200 km
750 km
2500 km
1 km
25 km
200 km
750 km
2500 km
OSV-5M/OSV-5M
vanilla
0.21
33.3
69.4
83.2
91.9
0.13
30.0
72.8
84.2
91.2
0.10
22.5
71.7
84.9
92.1
+1
0.26
33.4
69.4
83.2
91.9
0.12
30.3
72.8
84.2
91.2
0.09
22.7
71.8
84.9
92.1
+2
0.27
33.4
69.4
83.2
91.9
0.12
30.3
72.8
84.2
91.2
0.09
22.7
71.8
84.9
92.1
+3
0.28
33.4
69.4
83.2
91.9
0.12
30.3
72.8
84.2
91.2
0.09
22.7
71.8
84.9
92.1
Appendix
Table 7: Analytic refinement on geolocalization. XL models trained on each benchmark’s training set (IM2GPS3K on MP16-Pro). Every row decodes the same generated samples. We bold improvements over vanilla >0.5 .
GeoCLIP
SatCLIP
TTE
Decoder
< 1 km
GFLOPs
< 1 km
GFLOPs
< 1 km
GFLOPs
Base
67.91
0.076
99.65
0.076
99.59
0.076
+1
99.10
0.208
100.00
0.093
100.00
0.124
+2
99.67
0.341
100.00
0.110
100.00
0.172
+3
99.79
0.473
100.00
0.127
100.00
0.220
+4
99.85
0.605
100.00
0.144
100.00
0.267
Appendix
Table 8: Analytic refinement across location encoders. Fraction of 1M uniform land coordinates localized within 1 km, and decoder cost, as a function of LM refinement steps. The highlighted row is our chosen configuration.
OSV-5M
YFCC4k
IM2GPS3K
Readout
GeoS.
Mean
Ctry
Reg.
City
1 km
25 km
200 km
750 km
2500 km
1 km
25 km
200 km
750 km
2500 km
Decoder
4095
852
83.4
60.1
24.6
30.1
47.5
58.3
71.1
82.9
17.3
41.6
56.2
72.0
86.3
Training set gallery
4040
878
82.2
56.2
19.9
28.4
46.7
58.0
71.0
82.9
17.3
41.4
56.2
71.7
86.3
Uniform-land gallery
4067
861
82.3
55.9
17.5
0.0
22.8
55.4
70.1
82.5
0.0
17.4
54.0
71.0
85.9
Appendix
Table 9: Decoding versus gallery retrieval. We compare reading out generated embeddings using our LAE decoders with retrieving over training-set or uniform-land location galleries. Reported metrics use DIG-XL GeoCLIP model (OSV-5M uses the OSV-5M model; YFCC4k the YFCC-44M model; IM2GPS3K the MP16-Pro model) Best in bold .
Model
Width w
Depth D
MLP hidden
Params
B
768
12
3072
82M
L
1024
24
4096
284M
XL
1152
28
4608
417M
Appendix
Table 10: Diffusion model sizes. Widths and depths follow DiT ( Peebles & Xie, 2023 ) ; the MLP hidden size is 4w . Parameter counts are for a 512 -dimensional location embedding and a 2048 -dimensional image condition, and exclude the REPA head ( 2 – 3 M parameters).
Latent d
512 (GeoCLIP), standardized per dimension
Width w / blocks / MLP expansion
1152 / 28 / 4
Parameters
420M
Alignment
k=8 , λ=0.5 , λcos=0.5
Time schedule
Eq. 11 , (a,b)=(−5,5)
Condition dropout pdrop
0.1
Optimizer
AdamW, β=(0.9,0.999) , weight decay 0.05
Appendix
Table 11: Training configuration of the XL model.
OSV-5M
YFCC4k
Method
precision ↑
recall ↑
density ↑
coverage ↑
precision ↑
recall ↑
density ↑
coverage ↑
Uniform
0.29
0.98
0.21
0.21
0.59
0.99
0.38
0.22
vMF
0.598
0.982
0.499
0.446
0.667
0.993
0.542
0.599
vMFMix
0.513
0.980
0.422
0.358
0.626
0.988
0.474
0.498
Diff R3
0.822
0.916
0.752
0.568
0.938
0.959
0.959
0.837
FM R3
0.845
0.907
0.799
0.575
0.953
0.959
1.037
0.920
Appendix
Table 12: Generative metrics of the unconditional distribution. Precision, recall, density and coverage Naeem et al. (2020) of samples drawn from p(y∣∅) against the test-set locations of OSV-5M (210k) and YFCC4k (4.5k), following Dufour et al. (2025) . Best in bold , second italicized .
OSV-5M
YFCC4k
IM2GPS3K
Rule
GeoS.
Mean
Ctry
Reg.
City
1 km
25 km
200 km
750 km
2500 km
1 km
25 km
200 km
750 km
2500 km
GeoCLIP
Single sample
4095
851
83.4
60.1
24.6
30.1
47.5
58.3
71.1
82.9
17.3
41.6
56.2
72.0
86.3
Center
4081
767
82.6
58.7
24.2
26.1
41.5
54.8
69.3
82.1
16.4
38.5
55.1
72.2
86.3
Cluster mean
4142
790
84.5
61.1
25.4
28.5
48.5
61.9
72.6
83.3
17.5
42.0
58.1
73.3
86.7
Cluster latent mean
4141
790
84.5
61.2
25.5
30.3
49.0
61.8
72.6
83.3
17.9
42.5
58.1
73.3
86.7
Appendix
Table 13: Aggregating 16 posterior samples. Each test image is sampled 16 times with different seeds, and predictions are combined by the rule in each row. OSV-5M uses the OSV-5M models; YFCC4k the YFCC-44M models; IM2GPS3K the MP16-Pro models. Best rule per encoder in bold .
Im2GPS3k
YFCC4k
Method
1 km
25 km
200 km
750 km
2500 km
Med.
1 km
25 km
200 km
750 km
2500 km
Med.
GeoRanker
18.6
45.0
61.5
76.2
89.2
63.5
32.6
43.3
53.8
70.0
82.6
121.2
GeoRanker + prior
18.9
46.5
62.9
76.9
89.4
47.7
33.4
45.4
55.8
71.4
83.4
79.9
GeoRanker + prior + samples
19.6
47.4
63.5
77.9
89.6
38.0
33.6
45.5
56.4
72.2
84.1
75.4
GeoRanker + prior + samples + snap
20.1
47.6
64.1
78.1
89.6
36.0
34.0
45.5
56.8
71.9
83.8
73.7
Appendix
Table 14: DIG as a prior for GeoRanker. Accuracy (%) within distance thresholds and median error (km). GeoRanker and DIG-XL (GeoCLIP) are trained on MP16-Pro.
Figure 8: Geographic distribution of training sets.
Figure 9: Geographic distribution of test sets.
Distance (km)
Within (%)
Training set
Median
Mean
1 km
25 km
Test: OSV-5M
OSV-5M
4.32
15.7
0.1
86.4
Test: Im2GPS3k
MP16-Pro
0.081
1.03
86.5
99.0
YFCC-44M
0.022
0.82
92.0
99.1
Appendix
Table 15: Train–test geographic overlap. For every test location, the great-circle distance to its nearest training location. “Within” columns give the share of test locations with a training location inside that radius.
IM2GPS3k
YFCC4k
Method
Source
# Images
Source
# Images
PlaNet ( Weyand et al., 2016 )
Web
126M
Web
126M
CPlaNet ( Seo et al., 2018 )
Flickr
30.3M
Flickr
30.3M
Translocator ( Pramanick et al., 2022 )
MP-16
4.7M
MP-16
4.7M
GeoDecoder ( Clark et al., 2023 )
MP-16
4.7M
MP-16
4.7M
PIGEOTTO ( Haas et al., 2024 )
Flickr/Wikipedia
4M
Flickr/Wikipedia
4M
Appendix
Table 16: Reported training sources of baselines on IM2GPS3k and YFCC4k.
Figure 10: Estimated posterior distribution. We use DIG-XL (GeoCLIP) trained on OSV-5M.
Worldwide image geolocalization, which aims to predict the GPS coordinates of any image on Earth, remains challenging due to global visual diversity. Recent generative approaches based on Retrieval-Augmented Generation (RAG) and Large Multimodal Models (LMMs) leverage candidates retrieved from fixed databases for reasoning, but often struggle with scenes that are absent from the reference set. In this work, we propose GeoSearch, an open-world geolocation framework that integrates web-scale reverse image search into the RAG pipeline. GeoSearch augments LMM prompts with database-retrieved coordinates and textual evidence extracted from web pages. To mitigate noise from irrelevant content, we introduce a two-layer filtering mechanism consisting of image matching, followed by confidence-based gating. Experiments on standard benchmarks Im2GPS3k and YFCC4k demonstrate the superiority of GeoSearch under leakage-aware evaluation. Our code and data are publicly available to support reproducibility.
Tung-Duong Le-Duc, Hoang-Quoc Nguyen-Son, Minh-Son Dao
University of Science, VNU-HCM Ho Chi Minh City, Vietnam · National Institute of Information and Communications Technology Tokyo, Japan
Image geolocation aims to estimate where a photograph was taken from its visual content. At worldwide scale, this remains challenging because visual evidence is often ambiguous, diverse, and unevenly distributed. Prior work has typically treated geolocation of ordinary internet photos and street-view imagery as separate tasks, despite their complementary strengths: internet photos better match the appearance distribution of user-captured queries, while street-view imagery provides denser, geographically grounded coverage. We present Pinpoint, a retrieve-and-rerank architecture that combines both sources in a coarse-to-fine pipeline. A contrastive image-GPS embedder is trained on both user-uploaded Flickr photos and street-view imagery, learning a shared image-GPS embedding space that is used to retrieve candidate locations. An attention-based reranker then rescores retrieved candidates by combining candidate-level visual and GPS features with cross-source evidence from nearby locations to ground the prediction. Unlike recent prior work, Pinpoint does not rely on multimodal large-language models, making inference faster and more reproducible. Pinpoint achieves state-of-the-art results across all metrics on standard benchmarks for internet photos (IM2GPS3k and YFCC4k) and street-view imagery (OSV-5M).
Worldwide image geolocalization aims to predict precise GPS coordinates for images captured anywhere on Earth, which is challenging due to the large visual and geographic diversity. Recent methods mainly follow two paradigms: retrieval-based approaches that match queries against a reference database, and generation-based approaches that directly predict coordinates using Large Vision-Language Models (LVLMs). However, we observe distinct error profiles between them: retrieval excels at fine-grained instance matching, while generation offers robust semantic reasoning. This complementary heterogeneity suggests that no single paradigm is universally superior. To harness this potential, we propose GeoRouter, a dynamic routing framework that adaptively assigns each query to the optimal paradigm. GeoRouter leverages an LVLM backbone to analyze visual content and provide routing decisions. To optimize GeoRouter, we introduce a distance-aware preference objective that converts the distance gap between paradigms into a continuous supervision signal, explicitly reflecting relative performance differences. Furthermore, we construct GeoRouting, the first large-scale dataset tailored for training routing policies with independent paradigm predictions. Extensive experiments on IM2GPS3k and YFCC4k demonstrate that GeoRouter significantly outperforms state-of-the-art baselines.
Pengyue Jia, Derong Xu, Yingyi Zhang +5
Department of Data Science, City University of Hong Kong