{"id":"6db29bd3-37a7-4591-8758-e1c8ae079a02","arxiv_id":"2501.00895","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A new 10.5M-pair remote sensing dataset and a 1.3B diffusion model generate resolution-controlled satellite imagery from text, with large reported gains on the RSICD benchmark.","lead":"The paper introduces Git-10M, a 10.5-million-pair global remote sensing image and text dataset, and Text2Earth, a 1.3 billion parameter diffusion model that generates remote sensing images from text and resolution prompts. It reports large gains on a standard benchmark and adds capabilities such as unbounded scene stitching and image editing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on Table II, but Text2Earth is pretrained on 10.5M image-text pairs and LoRA-fine-tuned on RSICD while baselines are trained from scratch, with the guidance scale selected on the test set; this confound, not the architecture, may explain the reported gains.","rationale":"The reader's verdict is REJECT with moderate confidence, and my read does not change that verdict. The benchmark confound is real, concrete, and directly attached to the strongest claim in the abstract and introduction: that Text2Earth outperforms previous models by a large margin on RSICD. I agree with the reader's weakest assumption that Table II is not a fair measure of model quality. In addition, the test-set guidance-scale sweep and the Algorithm 2 inconsistency strengthen the case that the reported numbers are not yet trustworthy. I am not arguing that large-scale pretraining is illegitimate as a design choice—the Git-10M dataset is a genuine contribution—but the paper's own framing claims Text2Earth 'outperforms previous models,' which requires isolating the model contribution from the data-scale advantage. The proposed re-benchmark settles this by controlling pretraining in both directions: if the margin survives matched pretraining, the concern is answered; if it collapses, the central claim must be rewritten as a dataset-plus-system result and compared against closely related models such as MetaEarth and DiffusionSat. Therefore the appropriate verdict remains REJECT (UNCHANGED), not because the resource is worthless, but because the paper's quantitative claim is not supported as written.","tokens_in":23951,"tokens_out":6283,"duration_ms":60364,"concrete_test":"Run a matched-protocol re-benchmark on the same RSICD split and evaluation code. Arm A: train Text2Earth from scratch on RSICD only (no Git-10M pretraining, same architecture and hyperparameters) and evaluate with omega chosen on a validation split. Arm B: take the strongest baseline (e.g., CRS-Diff or DiffusionSat) and give it the same advantage—pretrain on Git-10M or initialize from a pretrained Stable Diffusion checkpoint, then LoRA-fine-tune on RSICD—with the same validation-based omega selection. Report FID and Cls-OA over at least three seeds. If Arm A collapses to baseline-level FID and Arm B matches or beats Text2Earth's Table II numbers, the headline improvement is largely a pretraining-scale artifact; if Text2Earth retains a large margin in both arms, the SOTA claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—state-of-the-art remote sensing text-to-image generation on RSICD (Table II, +26.23 FID, +20.95% Cls-OA)—rests on a comparison that is not apples-to-apples. Text2Earth is pretrained on Git-10M (10.5M image-text pairs) and then LoRA-fine-tuned on RSICD (Section V-D), while the baselines in Table II are trained from scratch at RSICD scale in their original protocols. The table therefore varies both method and pretraining data scale simultaneously. With a 1.3B-parameter diffusion model and orders of magnitude more training data, a large FID gain over smaller GANs trained on roughly 10k images is expected and does not by itself evidence that the proposed resolution guidance or dynamic condition adaptation improves on prior methods. The evaluation also selects the guidance scale omega=3.0 directly on the RSICD test set (Table III), with no validation split and no error bars; the reported 24.49 FID / 90.26% Cls-OA is thus an optimized selection, not a stable estimator. This concern is load-bearing because the +26.23 FID number is the paper's headline evidence for its main contribution. A secondary internal inconsistency appears in Algorithm 2: line 8 computes guided noise epsilon_g from (tau, rho) conditionals, but line 9 performs the DDPM update using epsilon_theta(z_t, t, tau) rather than epsilon_g, so the exact sampling procedure described cannot be reproduced as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Git-10M, a global remote sensing image-text dataset of 10.5 million pairs with resolution and geospatial metadata, and Text2Earth, a 1.3-billion-parameter latent diffusion model with resolution guidance and a dynamic condition adaptation strategy. The authors report state-of-the-art results on the RSICD benchmark (FID 24.49, zero-shot Cls-OA 90.26%, CLIP score 25.62, corresponding to a claimed +26.23 FID and +20.95% Cls-OA improvement over prior methods), and demonstrate additional capabilities in unbounded scene construction, image editing, and cross-modal generation. The paper also studies data augmentation and a CLIP-style pretraining use of Git-10M.","tokens_in":24250,"tokens_out":4108,"duration_ms":38727,"significance":"If the empirical claims hold, the dataset and model are valuable resources for remote sensing generation: Git-10M is an order of magnitude larger than prior remote sensing image-text datasets, and Text2Earth targets resolution-controllable, large-scale generation, which is a real gap. The paper also reports qualitative results across several downstream tasks and documents a concrete limitation (object-count control) in Section VI, which is a helpful sign of honest evaluation. The main weakness is that the headline quantitative claims rest on an evaluation protocol that does not support the attribution of the gains to the proposed method, and the sampling algorithm as written is internally inconsistent. These issues are load-bearing for the central SOTA claim, so the paper needs substantial revision before the claims can be accepted at face value.","major_comments":[{"comment":"The RSICD comparison is not apples-to-apples: Text2Earth is pretrained on Git-10M (10.5M image-text pairs) and then LoRA-fine-tuned on RSICD, while the baselines in Table II (AttnGAN, DF-GAN, Lafite, DALL-E, Txt2Img-MHN, RSDiff, CRS-Diff) are trained from scratch at RSICD scale in their original protocols. The comparison therefore varies both the method and the amount of training data, so the +26.23 FID and +20.95% Cls-OA improvements cannot be attributed to the resolution guidance or dynamic condition adaptation contributions. The authors should either add controlled baselines (e.g., a Stable Diffusion or latent diffusion model initialized from the same pretrained weights and fine-tuned on RSICD without the proposed components), or explicitly reframe the claim from 'our architectural contributions are superior' to 'our full system outperforms prior from-scratch methods,' with appropriate ablation support.","section":"Table II and Section V-D"},{"comment":"The guidance scale omega is selected directly on the RSICD test set: Table III reports FID and Cls-OA for omega from 1.5 to 7.0, and omega=3.0 is chosen because it gives a 'favourable trade-off' on that test set. No validation split is used, and no error bars or repeated-run statistics are reported. The headline FID 24.49 is therefore an optimized selection on the evaluation set, not an unbiased estimate of expected performance. The authors should use a held-out validation split for hyperparameter selection or, at minimum, report the sensitivity and variance (e.g., multiple seeds) so the reader can judge the stability of the claimed improvement.","section":"Table III and Figure 11"},{"comment":"The sampling procedure as written is internally inconsistent. Line 8 computes the guided noise epsilon_g = (1+omega)*epsilon_theta(z_t,t,tau,rho) - omega*epsilon_theta(z_t,t,tau_empty,rho_empty), but line 9 performs the DDPM update using epsilon_theta(z_t,t,tau) rather than epsilon_g. As written, the update ignores the resolution embedding rho and the classifier-free guidance combination entirely, so the algorithm cannot be reproduced as stated. The authors must correct the update to use epsilon_g and verify that the reported experiments used the corrected procedure; if the published numbers were produced with the line-9 form, the relationship between the algorithm and the results needs to be clarified.","section":"Algorithm 2, lines 8-9"},{"comment":"The comparison omits the most relevant recent foundation-model baselines. MetaEarth (reference [9], from the same research group) and DiffusionSat (reference [72]) are remote sensing generative foundation models trained on large-scale data, yet neither appears in Table II, and MetaEarth is not discussed anywhere in the body despite being a direct predecessor. If the paper claims to 'surpass previous models' for remote sensing text-to-image generation, the table should include these strong baselines or explain why they are not comparable. Without them, the SOTA claim is not fully supported even setting aside the pretraining confound.","section":"Table II and Related Work"}],"minor_comments":[{"comment":"The dataset size is stated inconsistently: the abstract reports 10.5 million image-text pairs, while Section V-A says '10 million global remote sensing image-text pairs.' Please reconcile these numbers.","section":"Section V-A and Abstract"},{"comment":"The condition-drop logic is inconsistent between text and resolution: for text, ctext==1 triggers dropping, while for resolution, cres==0 triggers dropping. This makes the pseudocode confusing and should be made uniform (e.g., both drop when the Bernoulli sample equals 1).","section":"Algorithm 1"},{"comment":"The 'Zero-shot Cls-OA' metric is not zero-shot in the usual sense: a ResNet-18 is trained on generated images labeled by text descriptions and then evaluated on real test images. Please clarify the terminology and cite the original source of this protocol more precisely, as it measures domain transfer rather than zero-shot recognition.","section":"Section V-C"},{"comment":"The x-axis labels are jumbled ('1.5 2.0 4.03.0 6.05.0 7.0'), which makes the figure difficult to read. Please fix the tick labels.","section":"Figure 11"},{"comment":"The phrase 'user-free text input' appears in several places and is ambiguous; it should likely read 'user-provided text input' or 'arbitrary text input.'","section":"Section I and IV"},{"comment":"Reference [9] (MetaEarth) is listed but not cited in the body text. Either cite it where relevant or remove it from the reference list to avoid an orphan citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising dataset contribution, but the empirical evaluation needs substantial strengthening. The most delicate issue is the omission of MetaEarth (a closely related foundation model from the same group) from the comparison table; this may draw scrutiny from reviewers and should be addressed head-on. The algorithm inconsistency is likely a typo, but it must be fixed because it currently undermines reproducibility. I would be open to accepting a revised version that adds controlled baselines, uses a validation split for hyperparameter selection, reports error bars, and includes the strongest existing foundation-model comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: Git-10M is the real news. A 10.5M image-text pair remote sensing dataset with resolution metadata is 5x the previous largest, and that's a solid contribution the community can build on. Text2Earth itself is a fairly standard latent diffusion model with a resolution embedding and classifier-free guidance; the resolution-control results look credible and the qualitative figures show real breadth. The paper is best read as a dataset + system paper, not as a rigorous method comparison.\n\nThe soft spot is Table II. Text2Earth is pretrained on Git-10M and LoRA-fine-tuned on RSICD, while the baselines are trained from scratch on RSICD. That comparison stacks method and data scale at the same time, so the +26.23 FID and +20.95% Cls-OA gains can't be attributed to the proposed mechanisms. On top of that, the guidance scale omega is chosen on the RSICD test set (Table III), no error bars are reported, and there's no direct comparison to MetaEarth or DiffusionSat. There's also an internal inconsistency in Algorithm 2: line 8 computes the guided noise epsilon_g, but line 9 uses the unguided epsilon_theta(z_t, t, tau) for the update, so the sampling procedure as written is unreproducible.\n\nThese are evaluation flaws, not fatal ones. The dataset claim holds up, the qualitative demonstrations are plausible, and the limitations section honestly acknowledges the model's object-count failures. With a controlled evaluation—same pretraining for baselines, a validation-based omega, and error bars—the SOTA claim might survive. As published, it doesn't.\n\nI'd send this to a serious referee: the dataset alone deserves review, and the model can be fixed. I'd cite the dataset in my own work. I just wouldn't repeat the RSICD numbers without checking what they actually controlled for.","headline":"Git-10M is a genuinely valuable dataset, but the Text2Earth SOTA claims rest on an apples-to-oranges benchmark comparison and a test-set-tuned guidance scale.","tokens_in":24868,"tokens_out":3031,"would_cite":true,"duration_ms":26188,"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":"Text2Earth pairs a 10.5-million-pair global satellite dataset with a 1.3-billion-parameter diffusion model to make text-driven remote sensing generation resolution-controllable and state of the art.","keywords":["remote sensing image generation","text-to-image diffusion","Git-10M dataset","resolution guidance","dynamic condition adaptation","unbounded scene construction","cross-modal image generation","foundation model"],"falsifier":"Re-running the RSICD comparison with all methods pretrained on the same Git-10M data, or with the guidance scale fixed on a validation split rather than the test set, would settle whether the +26.23 FID and +20.95 percentage-point Cls-OA gaps persist.","tokens_in":23679,"feed_emoji":"🛰️","tokens_out":7487,"duration_ms":64584,"temperature":0.7,"pith_summary":"The paper aims to establish that text-driven remote sensing image generation can move from small, scene-specific datasets and fixed-size outputs to a global-scale, resolution-controllable foundation-model capability. It contributes Git-10M, a corpus of 10.5 million remote sensing image-text pairs that is five times larger than the previous largest dataset and includes resolution metadata, and Text2Earth, a 1.3-billion-parameter latent diffusion model that injects a resolution embedding at every denoising step and uses a dynamic condition adaptation strategy. On the RSICD benchmark, Text2Earth reports a +26.23 improvement in Fréchet Inception Distance (FID) and a +20.95 percentage-point gain in zero-shot classification overall accuracy over prior models, while also demonstrating zero-shot generation, unbounded outpainting, image editing, and cross-modal synthesis. If these results hold, the work turns remote sensing text-to-image generation into a flexible tool for synthetic data, broad-area visualization, and multi-sensor imagery.","feed_headline":"Text2Earth beats prior satellite text-to-image models by 26 FID points","feed_subtitle":"A 1.3B-parameter diffusion model trained on 10.5M image-text pairs adds resolution control, outpainting, and editing","key_machinery":"The carrying mechanism is the conditioning stack of Text2Earth. A VAE compresses images into a compact latent space; a U-Net predicts noise in that space; text is injected through cross-attention from an OpenCLIP ViT-H encoder; and a resolution embedding module maps the requested ground-sample distance to a vector that is added to the timestep embedding at each denoising step. The dynamic condition adaptation strategy is the training-side counterpart: text and resolution conditions are randomly dropped during training, and at sampling the model blends a conditional noise prediction with an unconditional prediction through classifier-free guidance with a guidance scale. This setup is what lets one model produce resolution-controlled images, keep generating when a condition is absent, and be fine-tuned cheaply via LoRA for new datasets and modalities.","core_discovery":"The central claim, stated on its own terms, is that a single diffusion foundation model pretrained on enough globally distributed, resolution-annotated image-text pairs can generate geospatially plausible remote sensing imagery from free text, at a user-specified ground-sample distance, and can be adapted with small parameter-efficient changes to editing, outpainting, and cross-modal tasks. Text2Earth encodes images into a latent space, denoises them with a U-Net that attends to OpenCLIP text embeddings, and adds a resolution embedding to the timestep embedding in every denoising step. The dynamic condition adaptation strategy randomly drops text and resolution during training, then uses classifier-free guidance with null embeddings at sampling, so the model remains functional and controllable when conditions are missing or partial. On the RSICD benchmark the paper reports FID 24.49 and zero-shot Cls-OA 90.26%, which it presents as the strongest published numbers on that dataset, and it shows qualitative results for resolution-conditioned generation, unbounded scene stitching, cloud removal, local editing, text-driven SAR/NIR/PAN generation, and image-to-image translation.","pith_inferences":["A testable extension the paper does not run: checking whether the resolution embedding is continuous, so prompts like 1.3 meters per pixel, an unseen value, produce scale-consistent imagery, or whether control only works at the discrete resolutions shown.","The augmentation experiments only probe scene classification; if the same synthetic images helped detection or segmentation, where object positions and bounding boxes matter, the claim of a general data-augmentation engine would be stronger.","The cross-modal results suggest the shared latent prior might transfer to other geophysical modalities such as hyperspectral or LiDAR, but the paper only demonstrates SAR, NIR, PAN, RGB, and simulated foggy or low-resolution variants.","The unbounded outpainting demonstrations are qualitative; measuring seam consistency and long-range geospatial coherence would turn the observed flexibility into a quantitative claim."],"forward_implications":["On RSICD, Text2Earth reports FID 24.49 and zero-shot Cls-OA 90.26%, improvements of +26.23 and +20.95 percentage points over prior published results.","Text2Earth generates images at specified resolutions from the same prompt, such as 0.5, 1, and 2 meters per pixel for forest and storage-tank scenes, with scale-consistent detail.","Iterative outpainting with a fixed resolution builds unbounded scenes, demonstrated at 3500 by 1100 and 3700 by 1300 pixels, with smooth transitions between land-cover types.","With LoRA adaptation, the model produces text-driven RGB, PAN, NIR, and SAR images, and with a ControlNet-style module it performs PAN/RGB/NIR translation, super-resolution, and dehazing.","Synthetic images from Text2Earth improve downstream scene-classification accuracy for VGG-19, ResNet-18, ViT-B-16, and Swin-S, for example from 92.21% to 96.10% for Swin-S."],"supporting_citations":[{"why":"Supplies the RSICD benchmark dataset on which Text2Earth is evaluated and fine-tuned.","marker":"[17]"},{"why":"Supplies the latent diffusion architecture (VAE plus U-Net) that Text2Earth adapts for remote sensing.","marker":"[19]"},{"why":"Supplies the OpenCLIP ViT-H text encoder used to embed text for cross-attention.","marker":"[20]"},{"why":"Supplies classifier-free guidance, the basis for the dynamic condition adaptation sampling rule.","marker":"[83]"},{"why":"Supplies LoRA, the parameter-efficient fine-tuning method used on RSICD and for multi-modal generation.","marker":"[86]"},{"why":"Supplies the Txt2Img-MHN baseline and the zero-shot Cls-OA evaluation protocol used in the comparisons.","marker":"[8]"},{"why":"Supplies the CRS-Diff baseline, the previous best published model Text2Earth compares against on RSICD.","marker":"[73]"},{"why":"Supplies the language model used to generate the text annotations for Git-10M.","marker":"[82]"},{"why":"Supplies the ControlNet-style conditioning module used for image-to-image translation experiments.","marker":"[62]"}],"fun_headline_variants":["Text2Earth: 1.3B-param model turns text into global satellite imagery","10.5M image-text pairs power Text2Earth's remote sensing generation","Text2Earth cuts FID by 26 on RSICD with diffusion and resolution control","From text to satellite: Text2Earth sets new benchmark with 26-point FID gain","Global-scale text-to-image for remote sensing: Text2Earth on 10.5M pairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the benchmark comparison measures model quality fairly: Text2Earth is pretrained on 10.5 million pairs and LoRA-fine-tuned on RSICD, while baselines are trained from scratch on RSICD, and the guidance scale is selected using the RSICD test set.","fun_headline_variants_meta":{"raw":{"variants":["Text2Earth: 1.3B-param model turns text into global satellite imagery","10.5M image-text pairs power Text2Earth's remote sensing generation","Text2Earth cuts FID by 26 on RSICD with diffusion and resolution control","From text to satellite: Text2Earth sets new benchmark with 26-point FID gain","Global-scale text-to-image for remote sensing: Text2Earth on 10.5M pairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3534,"prompt_tokens":1122,"completion_tokens":2412,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":738,"completion_tokens_details":{"reasoning_tokens":2280}},"tokens_in":738,"tokens_out":2412,"duration_ms":16364,"temperature":1.0,"reasoning_tokens":2280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:40:30.547422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the RSICD comparison with all methods pretrained on the same Git-10M data, or with the guidance scale fixed on a validation split rather than the test set, would settle whether the +26.23 FID and +20.95 percentage-point Cls-OA gaps persist.","supporting_citations":[{"cited_title":"Exploring models and data for remote sensing image caption generation,","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion architecture (VAE plus U-Net) that Text2Earth adapts for remote sensing."},{"cited_title":"Adding conditional control to text-to-image di ffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the ControlNet-style conditioning module used for image-to-image translation experiments."}],"review_version":1}