{"id":"7e9b9ea9-f29b-4241-b631-7ba97d7f3d51","arxiv_id":"2504.19432","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EarthMapper uses geo-conditioned joint scale autoregression with key-point guidance to set a new state of the art on bidirectional satellite-map translation, supported by a new 302k-pair Chinese city dataset.","lead":"This paper presents EarthMapper, an autoregressive model that translates satellite images into maps and maps into satellite images in both directions, along with a large new dataset of 302,132 aligned pairs from 38 Chinese cities. The method adds geographic coordinates and a joint multi-scale generation process, and the authors report large gains over prior image-to-image translation models on standard metrics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on guidance scales tuned directly on the test set (Tables IV/V), so the reported margins over baselines are not established as expected performance.","rationale":"The paper's central claim is that EarthMapper sets a new state of the art in bidirectional satellite-map translation. For that claim to hold, EarthMapper must outperform the baselines under a fair and reproducible evaluation protocol. The most load-bearing weakness is not the internal conditioning mechanism but the evaluation protocol: the guidance scales in Tables IV and V are selected on the New York test set, and the chosen values are then used for the headline results in Table I. This means the reported FID, KID, SSIM, and PSNR are post-selection extrema rather than expected performance for a fixed model, and the baselines are not granted the same per-direction test-set tuning. The absence of error bars makes small margins, such as SSIM 0.6534 versus 0.6465, especially fragile. The reader's identified concern about key point force is legitimate and should be fixed or empirically defended, but the ablation evidence shows that KPF is not the sole source of the claimed gains; removing it still leaves FID 34.69 on CNSatMap, which remains below ControlVAR's 53.07. Thus the conditioning arithmetic is a secondary issue. The absence of released code and data also prevents independent verification, reinforcing the conditional verdict. Because the central claim may still survive under a corrected evaluation, a conditional verdict is appropriate rather than rejection, and the reader's CONDITIONAL verdict should remain unchanged.","tokens_in":20269,"tokens_out":4444,"duration_ms":49028,"concrete_test":"Split the New York dataset into a validation split and a held-out test split. For each guidance scale in Tables IV and V, run EarthMapper on the validation split, select the best scale per direction, freeze that configuration, then evaluate on the held-out test split. Repeat with at least three random seeds and report mean and standard deviation for FID, KID, SSIM, and PSNR. Perform the same validation-based tuning for the baseline methods. If EarthMapper still beats ControlVAR and BBDM by margins larger than the run-to-run spread, the SOTA claim survives; if the margins collapse or overlap within error, the headline overclaims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The SOTA claim is not established by the reported evaluation protocol. In Section V.D.2 (Tables IV and V), the guidance scale is swept on the New York test set and the best scale is chosen per direction: [8,8,8] for map-to-satellite (FID 36.54) and [2,2,2] for satellite-to-map (SSIM 0.6534). Table I then reports exactly these post-selection numbers as EarthMapper's performance. This is test-set hyperparameter selection: the headline metrics are the best of a small grid, not the performance of a fixed configuration, and the baselines were not given the same per-direction test-set tuning. Small reported advantages such as SSIM 0.6534 versus ControlVAR's 0.6465 and PSNR 25.04 versus 24.40 are also reported once with no error bars or multiple seeds, so they may lie within run-to-run noise. The key-point-force concern in the reader's verdict is real: Eq. (19) adds unordered codebook indices, which is not a principled conditioning operation. However, the ablation on CNSatMap shows that removing KPF leaves FID at 34.69, still competitive with ControlVAR at 53.07, so KPF is not the main load-bearing element. The decisive issue is that the claimed state-of-the-art margins have not been measured under an honest, fixed evaluation protocol with validation-based hyperparameter selection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents EarthMapper, a visual autoregressive framework for bidirectional satellite-map translation (BSMT). The method combines geographic coordinate embeddings, multi-scale hierarchical residual quantization, a geo-conditioned joint scale autoregression (GJSA) transformer, a semantic infusion (SI) loss using a frozen DINOv2 encoder, and an inference-time key point adaptive guidance (KPAG) scheme consisting of key point force (KPF) and complexity guidance (CG). The authors also introduce CNSatMap, a large-scale dataset of 302,132 aligned satellite-map pairs across 38 Chinese cities. Experiments on CNSatMap and the New York dataset report state-of-the-art FID, KID, SSIM, PSNR, RMSE, and LPIPS numbers, along with ablations and zero-shot demonstrations for in-painting, out-painting, and coordinate-conditional generation.","tokens_in":20544,"tokens_out":3570,"duration_ms":38119,"significance":"If confirmed, the work would be a meaningful advance in remote-sensing image translation: it is one of the first autoregressive frameworks to unify bidirectional satellite-map translation in a single training cycle, and the CNSatMap dataset is a potentially valuable benchmark resource. The paper contains useful design ingredients, and the ablations in Table III show monotone improvements from each added component. The zero-shot in-painting/out-painting and coordinate-conditional experiments are promising, and the promised public release of code and data is commendable. However, the headline state-of-the-art claim is currently not well supported because the best guidance scales are selected using the test set, and the mathematical description of the inference-time guidance has unresolved issues.","major_comments":[{"comment":"The reported headline numbers in Table I are obtained by selecting the guidance scale on the New York test set: Table IV picks [8,8,8] for map-to-satellite and Table V picks [2,2,2] for satellite-to-map, and these exact configurations are then reported as EarthMapper's performance in Table I. This is test-set hyperparameter selection, not evaluation of a fixed configuration. The baselines in Table I are not given the same per-direction test-set tuning, so the comparison is not apples-to-apples. Furthermore, all metrics are reported as single point estimates with no error bars, multiple seeds, or statistical significance tests; margins such as SSIM 0.6534 vs. 0.6465 for ControlVAR may lie within run-to-run noise. The central claim of state-of-the-art performance should be re-established with validation-based hyperparameter selection, repeated seeds, and a fair tuning protocol for baselines.","section":"V.D.2 and Tables IV/V vs. Table I"},{"comment":"The key point force mechanism adds the conditional image's quantized codebook index to the generated index at selected key points. Codebook indices are categorical and unordered, so adding them with clipping (Eq. (19)) is not a semantically meaningful operation in the token space. The paper states that this 'integrates the information from the key points into the generated image', but no formal argument or dedicated experiment shows that this operation biases the autoregressive sampler toward the intended conditional target; the semantics of 'key point' selection in Eq. (17) is likewise based only on the magnitude of a normalized index. Since KPF is a core component of the claimed controllable bidirectional translation mechanism, the authors should either provide a principled justification for index addition or replace it with an operation that provably maps conditional information into the generative distribution.","section":"IV.F, Eq. (19)"},{"comment":"Equation (23) defines the complexity-guided CFG as pθ(xi|x<i,c) = pθ(xi|x<i) + s(xi,ϕ)·∇pθ(xi|x<i) log pφ(c|xi). As written, this is not a normalized probability distribution, and the gradient notation ∇pθ(xi|x<i) log pφ(c|xi) is undefined: the gradient is not taken with respect to the probability mass function, and no practical algorithm is given for computing this term for discrete token probabilities. This is not a cosmetic issue: Table III shows that complexity guidance contributes a meaningful improvement (FID 31.57→29.89, SSIM 0.7054→0.7300), so the mechanism is load-bearing. The authors should provide a precise, implementable definition of this guidance step, including how the gradient is computed, how normalization is handled, and how α(ri), β(C(xi)) and γ are set in practice.","section":"IV.G, Eq. (23)"}],"minor_comments":[{"comment":"The text says 'the map portion of the generated paired image is substituted with a vector-quantized real map' but the subsequent paragraph notes that the architecture prevents direct information exchange between generated paired images; the relationship between these two statements should be clarified.","section":"IV.F, paragraph after Eq. (15)"},{"comment":"The claim that 'EarthMapper's Recall of 0.4890 (New York) ... exceeds ControlNet's 0.2110' is correct for New York, but the same sentence discusses CNSatMap where ControlNet's Precision (0.4886) exceeds EarthMapper's (0.4294); the text should acknowledge this asymmetry explicitly.","section":"V.C, paragraph 5"},{"comment":"The notation Q(f_m) in Eq. (2) is not defined until later, and Eq. (7) applies HRQ to each modality but the shared-codebook statement in the text should be made explicit in the equations.","section":"IV.C, Eq. (2) and Eq. (7)"},{"comment":"There are minor typos and grammatical infelicities, e.g., 'we redefine' should be 'we define', 'a AR-based' should be 'an AR-based', and the phrase 'the optimisation achieved during training' in Section IV.F should be reworded.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core weakness is not the idea but the evaluation and the formal specification of the inference guidance. The test-set tuning issue directly affects the validity of the stated state-of-the-art margins, and the ambiguous equations in Section IV.G should be resolved. If the authors can re-run the evaluation with a proper validation protocol and clarify or fix the guidance formulation, the paper could become a strong contribution. I would also ask the editor to ensure that the promised dataset and code release are verified during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read for you. The dataset is the real contribution here; the method is a competent adaptation of VAR/ControlVAR to satellite-map pairs, and the authors deserve credit for building CNSatMap. The SOTA claims, though, are not yet trustworthy.\n\nWhat's new: CNSatMap, 302k aligned satellite-map pairs across 38 Chinese cities, is a genuine resource for remote sensing and cartography. Also, jointly modeling satellite and map token maps in one next-scale autoregressive training cycle, conditioned on geo-coordinates, is a sensible extension not in ControlVAR or CAR. The ablation is informative: each added component gives a monotone improvement, and even the version without the questionable KPAG modules reaches FID 34.69 on CNSatMap, far below ControlVAR's 53.07. So the qualitative claim that the base architecture is strong is likely right.\n\nThe soft spots are real and non-trivial. First, the guidance-scale ablation in Tables IV/V is done on the New York test set, and the 'best' scales are then used for the headline numbers in Table I. That is test-set hyperparameter selection, and baselines were not given the same tuning. No error bars or multiple seeds either, so the small SSIM/PSNR margins over ControlVAR may be noise. Second, Eq. (19), the key-point force, adds unordered codebook indices; that operation has no semantic justification. It empirically helps, so maybe it works as a regularizer, but the paper does not show that. Third, Eq. (23) presents complexity guidance in a form that is not a valid probability update (a probability plus a gradient term). The prose describes CFG, but the equation does not match CFG as usually written. This needs fixing or a clearer explanation.\n\nAlso, code and data are promised but not available; for a paper whose central asset is the dataset, release matters.\n\nWho this is for: anyone working on satellite-map translation, geospatial generation, or controllable autoregressive models. The dataset makes it worth a look even if the method were weaker.\n\nBottom line: the core idea and dataset deserve a serious referee, and I would accept a revised version that releases the data, fixes the evaluation protocol, and cleanly reformulates KPF and CG. As it stands, treat the SOTA numbers as optimistic.","headline":"CNSatMap is the real contribution; the SOTA margins are compromised by test-set tuning and a couple of questionable formulas, but the base architecture is strong enough that the paper deserves a serious referee.","tokens_in":21110,"tokens_out":3652,"would_cite":true,"duration_ms":36847,"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":"EarthMapper claims one autoregressive model translates satellite images to maps and back, outperforming GAN, diffusion, and prior AR baselines.","keywords":["bidirectional satellite-map translation","visual autoregressive model","geo-conditioned joint scale autoregression","semantic infusion","key point adaptive guidance","CNSatMap dataset","controllable image generation","remote sensing"],"falsifier":"Run the map-to-satellite test set with the key-point index addition replaced by the identity mapping, keeping all other settings fixed, and recompute FID and KID; if the metrics do not materially worsen, the index-addition step is not the source of the reported conditional fidelity.","tokens_in":19994,"feed_emoji":"🛰️","tokens_out":8852,"duration_ms":80185,"temperature":0.7,"pith_summary":"EarthMapper claims that one visual autoregressive model can translate satellite imagery into maps and maps into satellite imagery, and that this single model outperforms the compared GAN, diffusion, and prior autoregressive baselines in both translation directions. The paper attributes the gain to three mechanisms: geographic-coordinate embeddings that anchor generation to a place, a joint-scale autoregressive training objective that unifies both directions in one training cycle, and inference-time key-point guidance that balances fidelity with diversity. A semantic-infusion loss aligns generated features with a pretrained vision model's features during training. The paper also contributes CNSatMap, a dataset of 302,132 aligned satellite-map pairs across 38 Chinese cities, and shows the same model handles zero-shot in-painting, out-painting, and coordinate-only generation. The potential payoff is automated cartography, rapid urban simulation, and augmentation of satellite datasets with rare scenes.","feed_headline":"One model translates maps to satellites and back","feed_subtitle":"It beats GAN and diffusion baselines, with map-to-satellite FID 29.89 on a new 302k-pair dataset.","key_machinery":"The load-bearing mechanism is the geo-conditioned joint scale autoregression (GJSA): both modalities are tokenized into multi-scale discrete token maps by hierarchical residual quantization, the satellite and map tokens are paired at each scale, and a transformer predicts each scale's token pair conditioned on all previous scales plus a sinusoidal embedding of latitude and longitude. That joint next-scale prediction gives the model a single training objective for both directions. At inference, the source modality's token maps are inserted into the joint sequence, and the key point force adds selected quantized indices of the condition to the sampled indices at high-importance positions; complexity guidance then modulates the conditioning strength as a function of resolution and image complexity. During training, a semantic infusion loss aligns the autoregressive hidden states with features from a frozen pretrained vision encoder, which the paper argues enforces feature-level consistency between generated and real images.","core_discovery":"The paper's central claim is that bidirectional satellite-map translation can be unified in a single multi-scale autoregressive framework, conditioned on geographic coordinates and on the source modality, so that map-to-satellite and satellite-to-map generation share one set of learned parameters. On the New York dataset it reports FID 36.54 and KID 0.99 for map-to-satellite translation, and SSIM 0.6534 with PSNR 25.04 for satellite-to-map; on the new CNSatMap dataset it reports FID 29.89 and KID 2.06 for map-to-satellite, and SSIM 0.7300 with PSNR 26.88 for satellite-to-map. In the paper's comparisons, these numbers beat the GAN, latent-diffusion, and autoregressive baselines included in the study on almost every reported metric, and the paper shows the same model producing coherent results for zero-shot in-painting, out-painting, and coordinate-only generation. The bidirectional capability follows from pairing the two modalities' token maps at each scale and treating a full scale's pair as one autoregressive step.","pith_inferences":["A testable extension is to replace the key-point index addition with a learned attention-based conditioner: codebook indices are unordered labels, so index arithmetic is not semantically grounded, and a learned conditioner that matches or beats the reported metrics would show the index addition is not what carries conditional control.","The cross-dataset results, CNSatMap-to-New York generalizing better than New York-to-CNSatMap, suggest that scale and geographic diversity of the training corpus, not the architecture alone, drive transfer; this points to a practical recipe of pretraining on large multi-city datasets before fine-tuning on a target region.","Coordinate-only generation opens a quantitative test of the coordinate embedding itself: generating at held-out coordinates with known ground truth would separate how much the output is determined by location versus by learned visual priors, a separation the current aggregate metrics do not provide."],"forward_implications":["If the reported numbers hold, EarthMapper establishes a new published benchmark on both the New York and CNSatMap test sets, with map-to-satellite FID roughly half that of the strongest prior autoregressive baseline.","Because one model handles both directions, deployment for mapping and simulation pipelines could use a single checkpoint instead of separate translation networks for each direction.","The CNSatMap dataset becomes a large-scale shared resource: 302,132 aligned pairs across 38 Chinese cities could support training and evaluation of future cross-modal geographic models beyond this architecture.","The geographic-coordinate embedding may enable generating imagery for arbitrary locations without retraining, which would matter for scenario simulation after disasters and for augmenting satellite datasets with rare scene types."],"supporting_citations":[{"why":"Introduces the next-scale autoregressive prediction scheme that EarthMapper's joint scale autoregression is built on.","marker":"[11]"},{"why":"Supplies the Pix2Pix baseline and the New York satellite-map dataset used for evaluation.","marker":"[5]"},{"why":"The controllable autoregressive baseline EarthMapper extends with geographic conditioning and compares against.","marker":"[12]"},{"why":"Represents the diffusion-based conditional-control baseline (ControlNet) in the comparisons.","marker":"[14]"},{"why":"Provides residual quantization, the basis of EarthMapper's hierarchical residual quantization tokenizer.","marker":"[41]"},{"why":"Defines the FID metric used for map-to-satellite evaluation.","marker":"[47]"},{"why":"Defines the KID metric used for distribution-distance evaluation.","marker":"[48]"},{"why":"Supplies the frozen pretrained vision encoder whose features drive the semantic infusion loss.","marker":"[51]"}],"fun_headline_variants":["Satellite-map translation, one autoregressive model, both directions","EarthMapper: bidirectional satellite-map translation in one model","Bidirectional satellite-map translation with one autoregressive model","Geographic coordinates anchor satellite-map translation in both directions","One autoregressive model translates satellite to map and back"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that adding the conditional image's quantized codebook indices to the generated image's indices at selected key points steers the autoregressive sampler toward the intended target, an assumption that is not proven and is not guaranteed because codebook indices are unordered labels.","fun_headline_variants_meta":{"raw":{"variants":["Satellite-map translation, one autoregressive model, both directions","EarthMapper: bidirectional satellite-map translation in one model","Bidirectional satellite-map translation with one autoregressive model","Geographic coordinates anchor satellite-map translation in both directions","One autoregressive model translates satellite to map and back"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001039,"raw_usage":{"total_tokens":4439,"prompt_tokens":1077,"completion_tokens":3362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":3284}},"tokens_in":693,"tokens_out":3362,"duration_ms":21893,"temperature":1.0,"reasoning_tokens":3284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:52:37.252705+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the map-to-satellite test set with the key-point index addition replaced by the identity mapping, keeping all other settings fixed, and recompute FID and KID; if the metrics do not materially worsen, the index-addition step is not the source of the reported conditional fidelity.","supporting_citations":[{"cited_title":"Visual autoregressive modeling: Scalable image generation via next-scale prediction,","cited_arxiv_id":null,"evidence_quote":"Introduces the next-scale autoregressive prediction scheme that EarthMapper's joint scale autoregression is built on."}],"review_version":1}