{"id":"c2731e1e-4579-4cf2-abe8-0233b670e6f6","arxiv_id":"2412.06781","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A generative model using diffusion and Riemannian flow matching on the sphere achieves state-of-the-art visual geolocation and outputs full probability maps over possible locations.","lead":"A computer vision team built a generative model that locates where a photo was taken by denoising random points on a globe, instead of predicting a single coordinate. The model also outputs a probability map over all possible locations, which helps when an image contains ambiguous cues like a beach or an animal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DDIM sampling loop as written (Eqs. 3–4) is an identity: substituting Eq. 4 into Eq. 3 gives x_{t-dt}=x_t, so the described inference cannot produce the reported accuracy.","rationale":"The paper's two headline contributions are state-of-the-art geolocation accuracy and calibrated predictive distributions. The reader's weakest assumption concerns the appendix density derivation (Prop. 2 sign, Eq. W missing 1/sqrt(κ)); those are real and I agree they need correction. However, they primarily affect the probabilistic extension and the Diffusion-R3 NLL row, not the accuracy results in Table 1. The cancellation in Eqs. (3)–(4) is upstream of both contributions: if the equations are read literally, the model cannot move the sampled coordinate, so the reported accuracy is impossible. Since the central claim is that a denoising process operating on coordinates yields state-of-the-art results, this internal inconsistency is the single most load-bearing concern. It is likely a typo—replacing κ(t) with κ(t−dt) in Eq. (3) restores a valid DDIM step—but until the authors correct the equations or release code, the manuscript as written does not support its central claim. I therefore keep a conditional verdict, but conditional on fixing and verifying the DDIM update as well as the appendix density formulas.","tokens_in":18326,"tokens_out":15165,"duration_ms":160725,"concrete_test":"Symbolically substitute Eq. (4) into Eq. (3) with arbitrary ψ: the right-hand side simplifies to x_t, proving the loop is a no-op. Then implement the update exactly as printed (using κ(t) in both terms) on an OSV-5M subset and measure GeoScore; if it is near the Uniform baseline, the equations cannot be the ones that produced Table 1. If the authors' code instead uses κ(t−dt), the paper must be corrected and the experiments re-verified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the core inference procedure, Eqs. (3) and (4) are mutually inconsistent. Substituting the expression for x̂_t from Eq. (4) into Eq. (3) yields x_{t-dt} = sqrt(1−κ(t))·(x_t − sqrt(κ(t))ψ(xt|c))/sqrt(1−κ(t)) + sqrt(κ(t))ψ(xt|c) = x_t, identically, for any network output ψ. As printed, the DDIM loop therefore does not refine the initial noise; after projection the output is just a random point on S2, which would score at chance level on all three benchmarks. The reported state-of-the-art numbers in Table 1 cannot follow from the published algorithm. The likely intended update uses κ(t−dt) in place of κ(t) in Eq. (3) (standard DDIM indexing), but the manuscript does not state this and no code is available. This is a more fundamental defect than the appendix density-formula sign and normalization issues: it affects the central geolocation accuracy claim, not only the probabilistic extension.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative approach to global visual geolocation. Instead of predicting a single point, the model denoises random locations conditioned on image embeddings, using Euclidean diffusion, Euclidean flow matching, or Riemannian flow matching on the sphere S^2. The authors claim state-of-the-art accuracy on OpenStreetView-5M, YFCC-100M, and iNat21, and introduce a probabilistic variant in which the model outputs a full distribution over locations via an ODE-based density computation. They also propose new metrics and baselines for probabilistic geolocation and show qualitative localizability results.","tokens_in":18530,"tokens_out":4619,"duration_ms":48101,"significance":"If the claims hold, this is a substantial contribution: it is the first generative treatment of global visual geolocation, it provides calibrated spatial distributions in addition to point predictions, and it demonstrates strong empirical results on three large benchmarks. The paper also adapts flow-matching density estimation to a compact manifold and introduces a new evaluation protocol for probabilistic geolocation, which could be useful for future work. The experimental comparison is broad, and the inclusion of non-parametric generative baselines (vMF, vMF mixtures) is appropriate.","major_comments":[{"comment":"The sampling procedure as printed is an identity. Substituting Eq. (4) into Eq. (3) gives x_{t-dt} = sqrt(1-kappa(t)) * [1/sqrt(1-kappa(t)) (x_t - sqrt(kappa(t)) psi(x_t|c))] + sqrt(kappa(t)) psi(x_t|c) = x_t, identically for any network output. The iterative denoising loop therefore does not modify the initial noise, and after projection the result would be a uniform random point on S^2, which cannot produce the accuracies reported in Table 1. The likely intended update uses kappa(t-dt) in Eq. (3) (standard DDIM indexing), but the manuscript does not state this and no code is available for verification. This is a load-bearing defect in the central inference algorithm and must be corrected and re-verified.","section":"Sec. 3.1, Eqs. (3) and (4)"},{"comment":"The 'corrected' Proposition 2 contains a sign error. Its statement claims log p(y|c) = log p_epsilon(x(1)|c) + f(1), while the proof in the same appendix (Eq. (O)) derives log p(x_0|c) = log p(x(1)|c) - f(1), which is consistent with Proposition 1 in the main text. The two are mutually contradictory. Additionally, Eq. (W) omits the 1/sqrt(kappa(t)) normalization when substituting the learned noise for the score: the correct velocity for the probability-flow ODE is -1/2 beta(t) (x - eps_theta / sqrt(kappa(t))), not -1/2 beta(t) (x - eps_theta). This affects the diffusion-model density estimates, so the DiffR3 NLL values and localizability numbers in Table 2 are unreliable as printed.","section":"Appendix D, Proposition 2 and Eq. (W)"},{"comment":"The NLL comparison in Table 2 mixes models defined in different metric spaces (R^3 versus S^2). The paper notes that these are not directly comparable, but then relies on this table to support the probabilistic-geolocation claim. Since the NLL of the R^3 models depends on an unnormalized density in the embedding space, the reported numerical differences (e.g., Diffusion R^3 NLL 0.58 vs. RFlowMatch S^2 NLL -1.51 on OSV-5M) should not be interpreted as relative quality without a common normalization; at minimum, the R^3 densities must be marginalized onto the sphere consistently before comparison.","section":"Sec. 4.2 / Table 2"}],"minor_comments":[{"comment":"The ablation text says the standard sigmoid scheduler 'increases the geoscore' relative to the proposed scheduler, and the table confirms this (3767.21 vs. 3746.79), yet the main text and Fig. 4 argue that the skewed scheduler gives better results. This contradiction should be clarified: either the scheduler choice is justified by density quality alone, or the numbers are misreported.","section":"Table A / Sec. A"},{"comment":"The abstract states 'Codes and models are available here' but no URL is given in the manuscript. Please provide a working link or state that code will be released upon publication.","section":"Abstract / Section 3.4"},{"comment":"The row for 'Uniform' under OSV-5M contains an excessive number of numerical entries (six values for five columns), making the table hard to read; please fix the alignment.","section":"Table 1"},{"comment":"There is a typo: 'New Zeland' should be 'New Zealand'.","section":"Sec. 4.2"},{"comment":"The phrase 'as measurs by the generative metrics' contains a typo; it should read 'as measured by the generative metrics'.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially strong, but the published inference equations (Eqs. 3–4) are self-cancelling, and the density-estimation appendix contains a sign error and a missing normalization factor. These are fixable within the scope of a revision, but they are load-bearing for the two central claims (geolocation accuracy and probabilistic prediction). I would recommend the editor ask for a corrected manuscript with a proper DDIM update, corrected density formulas, and recomputed tables, plus a code release to verify the results. I do not see a circularity problem: the evaluation is standard and the self-citations are to benchmark and method papers, not to the paper's own results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the real news: this is the first generative treatment of global visual geolocation, and it works. The authors train diffusion and flow-matching models to denoise points on the Earth's surface (Riemannian flow matching on S2), introducing probabilistic geolocation with NLL, localizability, and adapted precision/recall metrics. On OSV-5M, YFCC-4k, and iNat21, RFM-S2 beats strong retrieval and regression baselines, including the previous SOTA on OSV-5M. The writing is clear and the empirical work is careful: shared backbone, same training budget for baselines, and a sensible set of generative baselines (vMF and mixtures). The localizability examples (NFL stadium, beach) are compelling and demonstrate the value of the probabilistic framing.\n\nThe soft spots are real but appear to be presentation errors rather than load-bearing flaws. The biggest is the DDIM update. As printed, Eq. (4) substituted into Eq. (3) gives x_{t-dt}=x_t identically, so the diffusion inference does nothing. This is almost certainly a typo—standard DDIM uses κ(t-dt) on the right-hand side—but as published the DiffR3 results in Table 1 cannot be reproduced. The stress-test worry that this undermines the whole accuracy claim is overstated: the headline results are carried by RFM-S2 (and FM-R3), which use the ODE integrators in Eqs. (8)/(12), not this update. But the paper still reports DiffR3 as part of the contribution, so the error must be fixed and code released.\n\nSecond, the density estimation appendix has a sign inconsistency: Prop. 2 states log p = log p_ε + f(1) and the integration step implies that, but Eq. (O) inside the proof flips the sign. Also, the diffusion velocity in Eq. (W) drops a 1/sqrt(κ(t)) factor in the score term. Both look like typos, but the NLL numbers in Table 2 for the diffusion model and the R3 methods depend on this formula, so they should be re-verified.\n\nThird, the scheduler ablation contradicts the main text: the paper says it uses a skewed sigmoid (α=-3, β=7), but the ablation shows the standard sigmoid gives GeoScore 3767 on OSV-5M—exactly the value in Table 1—while the skewed one gives 3747. So the reported result seems to come from the standard scheduler, not the proposed one. This needs clarification.\n\nWho is this for? Anyone working on geolocation, uncertainty in CV, or spherical generative models. It deserves peer review, but with corrections requested: fix the DDIM equation, reconcile the scheduler, clean up the density appendix, and release code. I'd cite it once the typos are ironed out.","headline":"First generative geolocation on the sphere with strong results, but the published inference equations and appendix need corrections before the probabilistic claims are reliable.","tokens_in":19104,"tokens_out":4679,"would_cite":true,"duration_ms":43783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generative model locates images by denoising random points on the Earth's surface.","keywords":["visual geolocation","generative model","diffusion models","flow matching","Riemannian manifolds","probabilistic prediction","localizability","sphere S^2"],"falsifier":"Train the same flow-matching model on a synthetic spherical dataset with a known closed-form density, such as a mixture of von Mises distributions, compute likelihoods both by the paper's ODE formula and by the closed form, and check whether the NLL values match; a sign or normalization error would show up directly in this comparison.","tokens_in":18082,"feed_emoji":"🌍","tokens_out":4337,"duration_ms":44474,"temperature":0.7,"pith_summary":"This paper proposes replacing deterministic visual geolocation with a generative process: start from random points on the globe and iteratively denoise them, conditioned on a frozen image embedding, until they converge on plausible capture locations. Because the process runs on the sphere, the same model yields both a point estimate and a full conditional distribution over every location on Earth. The authors report higher accuracy than prior methods on OpenStreetView-5M, YFCC-100M, and iNat21, and they introduce the task of probabilistic visual geolocation with new metrics and baselines. The reason to care is that ambiguity is intrinsic to geolocation, so a density over locations is a more honest and more useful output than a single point.","feed_headline":"Denoising random globe points finds where an image was shot","feed_subtitle":"One generative model returns both a location and a probability over all possible spots, on three global benchmarks.","key_machinery":"The core object is a conditional denoising network ψ that takes as input a noisy coordinate xt, the noise level κ(t), and a frozen image embedding, and learns to predict the noise (diffusion) or the velocity field (flow matching) that moves points toward the true location. The Riemannian flow matching variant operates entirely on the sphere S2 using exponential and logarithmic maps, so every intermediate point lies on the Earth's surface. The probability-density claim is carried by an ODE system that integrates the negative divergence of the learned velocity field along the trajectory, turning the generative model into a likelihood estimator.","core_discovery":"The central claim is that visual geolocation can be reframed as a generative modeling problem: a network is trained to reverse a noising process that starts from an image's true location and corrupts it into pure noise, so that at inference time the network denoises a random starting point into a likely location. The authors implement this with three variants, diffusion in R3, flow matching in R3, and Riemannian flow matching directly on the sphere S2, and report that the spherical variant performs best, improving GeoScore by hundreds of points over the previous hybrid approach on OpenStreetView-5M while also giving lower average distance and higher country-level accuracy. Beyond the point prediction, the paper claims the learned velocity field can be integrated with a divergence term to compute log p(y|c) for any location y, which yields a calibrated probability distribution over the globe and a quantification of an image's localizability. The paper introduces probabilistic visual geolocation as a task, proposes NLL, localizability, precision, recall, density, and coverage as metrics, and reports that the diffusion and flow-matching models beat von Mises-Fisher baselines on these metrics.","pith_inferences":["Editorial inference: if the density estimates are truly calibrated, the model could serve as a spatial prior that is re-weighted by additional cues such as known country, language, or time of day, a combination the paper does not explore.","Editorial inference: the same spherical flow-matching setup should transfer to other spherical regression tasks, such as camera pose estimation or planet-scale remote sensing, where injecting Euclidean noise is geometrically inappropriate; the paper only tests it on geolocation.","Editorial inference: the distribution over locations could be used as a proposal for retrieval, sampling multiple trajectories and ranking database images by the computed density, potentially improving the fine-grained accuracy where retrieval methods currently beat generative ones.","Editorial inference: the reported NLL and localizability numbers depend on the appendix's density formula, and the appendix contains a sign discrepancy between its corrected proposition and its proof, so an independent check on a synthetic spherical density would be the cheapest way to verify the probabilistic results before building on them."],"forward_implications":["For any new image, the model can return both a single location estimate and a full probability density, so downstream systems can use the distribution to express confidence or to list multiple plausible locations instead of committing to one point.","The guidance scale acts as a dial between localization accuracy and distribution fidelity: higher guidance sharpens the predicted density and improves geolocation scores, while lower guidance preserves the full ambiguity of the image.","Because the denoising field is defined on the sphere rather than in Euclidean space, the approach respects the Earth's geometry and avoids the projection step that the Euclidean variants require; the authors report that the spherical variant consistently outperforms the Euclidean ones.","On YFCC-100M, extending training from one million to ten million iterations yields steady accuracy gains, suggesting the model continues to improve with more compute rather than saturating quickly.","The predicted distributions can distinguish images that are pinpointable to meter-level accuracy, such as the Eiffel Tower, from images with only continental information, such as a featureless beach, which is directly useful for applications like forensic verification and archival prioritization."],"supporting_citations":[{"why":"Supplies the Riemannian flow matching framework on general geometries that the spherical denoising process is built on.","marker":"[9]"},{"why":"Provides the flow matching training objective and the density-estimation derivation that the paper extends to geolocation.","marker":"[45]"},{"why":"Provides the OpenStreetView-5M dataset and the deterministic hybrid baseline that the paper compares against.","marker":"[2]"},{"why":"Provides the iNat21 benchmark used for evaluation of geolocation accuracy and distribution quality.","marker":"[74]"},{"why":"Provides the YFCC-100M dataset used for large-scale evaluation.","marker":"[1]"},{"why":"Introduces the YFCC4k evaluation subset and retrieval-based baselines that the paper reports against.","marker":"[76]"},{"why":"Provides the von Mises-Fisher mixture probabilistic baseline that the paper's generative models are benchmarked against.","marker":"[36]"},{"why":"Supplies the DDIM iterative sampling procedure used for inference in the diffusion variant.","marker":"[70]"},{"why":"Provides the relation between the diffusion SDE and the velocity field, used in the density prediction for the diffusion model.","marker":"[71]"}],"fun_headline_variants":["Generative geolocation: from noise to world coordinates","Spherical diffusion pinpoints any photo on Earth","Denoise a globe: one model gives location and odds","Probabilistic geolocation: a distribution over every spot","From random globe points to calibrated location guesses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole probabilistic side of the paper rests on a formula saying that solving one differential equation along the model's trajectories gives true likelihoods; the appendix's corrected statement disagrees with its own proof about a sign, and the diffusion velocity formula omits a normalization factor, so if those are not harmless errors the reported distribution numbers collapse.","fun_headline_variants_meta":{"raw":{"variants":["Generative geolocation: from noise to world coordinates","Spherical diffusion pinpoints any photo on Earth","Denoise a globe: one model gives location and odds","Probabilistic geolocation: a distribution over every spot","From random globe points to calibrated location guesses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1375,"prompt_tokens":944,"completion_tokens":431,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":354}},"tokens_in":560,"tokens_out":431,"duration_ms":5038,"temperature":1.0,"reasoning_tokens":354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:18:53.821623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same flow-matching model on a synthetic spherical dataset with a known closed-form density, such as a mixture of von Mises distributions, compute likelihoods both by the paper's ODE formula and by the closed form, and check whether the NLL values match; a sign or normalization error would show up directly in this comparison.","supporting_citations":[{"cited_title":"In: ICLR (2024) 2, 4, 7","cited_arxiv_id":null,"evidence_quote":"Supplies the Riemannian flow matching framework on general geometries that the spherical denoising process is built on."},{"cited_title":"In: The Eleventh International Conference on Learning Representations (2024) 2, 3, 5","cited_arxiv_id":null,"evidence_quote":"Provides the flow matching training objective and the density-estimation derivation that the paper extends to geolocation."},{"cited_title":"In: CVPR (2021) 1, 2, 5, 6, 8","cited_arxiv_id":null,"evidence_quote":"Provides the iNat21 benchmark used for evaluation of geolocation accuracy and distribution quality."},{"cited_title":"In: ICCV (2017) 1, 2, 6","cited_arxiv_id":null,"evidence_quote":"Introduces the YFCC4k evaluation subset and retrieval-based baselines that the paper reports against."},{"cited_title":"In: MLKDD (2020) 2, 6","cited_arxiv_id":null,"evidence_quote":"Provides the von Mises-Fisher mixture probabilistic baseline that the paper's generative models are benchmarked against."},{"cited_title":"In: ICLR (2021) 3, 15","cited_arxiv_id":null,"evidence_quote":"Supplies the DDIM iterative sampling procedure used for inference in the diffusion variant."},{"cited_title":"In: ICLR (2021) 2, 3, 15","cited_arxiv_id":null,"evidence_quote":"Provides the relation between the diffusion SDE and the velocity field, used in the density prediction for the diffusion model."}],"review_version":1}