CrossTimeEdit: A Decade-Spanning Cross-View Dataset and Reward-Guided Editing for Historical Street-View Generation

1 INTRODUCTION

• Existing datasets do not jointly provide temporal cross-view evidence and editing supervision for historical street-view recovery.Noise-based cross-view generation may alter persistent buildings and roads because satellite imagery sparsely constrains facades, textures, and occluded structures.Source: §1 · p. 2
• CrossTimeEdit uses recent street views to constrain viewpoint and unchanged appearance, while temporal satellite differences guide edits to changed regions.It starts from FLUX.2 [Klein] 4B and is trained with SFT followed by online RL.Source: §1 · p. 2
• VIGOR-his provides 43,653 decade-spanning cross-view quadruplets across 11 cities on three continents, with change descriptions and satellite-grounded local editing instructions.Its automated pipeline includes spatial pairing, consistency screening, change classification, instruction generation, and validation.Source: §1 · p. 2
• 17.12% improvement in overall performance over the pretrained backbone is reported for CrossTimeEdit, which also leads evaluated open-source models in editing and cross-view perceptual comparisons.The model is trained through SFT and GDPO-normalized Flow-GRPO.Source: §1 · p. 3

2.1 CROSS-VIEW AND TEMPORAL STREET-VIEW DATASETS

• Existing cross-view and temporal street-view datasets primarily support localization, retrieval, or temporal analysis, and do not jointly provide location-matched earlier and recent street–satellite views with local editing instructions.CityPulse includes street-view time series and change labels but lacks corresponding satellite pairs.Source: §2.1 · p. 3

2.2 IMAGE EDITING AND GENERATIVE POST-TRAINING

• Generic editing and preference-optimization methods may not suit street-view images with substantial distortion and panoramic geometry.The reviewed methods include instruction-based editing, spatial conditioning, image–text evaluation, and group-relative optimization for flow models.Source: §2.2 · p. 3

2.3 SATELLITE-TO-STREET GENERATION

• Satellite-to-street methods infer appearance predominantly from satellite imagery, leaving persistent street content and temporal change localization underconstrained.The passage notes this may lead to lower-quality street-view outputs.Source: §2.3 · p. 3

3 VIGOR-HIS

• VIGOR-his contains 43,653 location-level cross-view quadruplets across 11 cities on three continents, with local editing instructions for changed samples.The dataset is intended to support location-level urban spatio-temporal analysis and historical street-view generation.Source: §3 · p. 3

3.1 DATA COLLECTION AND PRE-PROCESSING

• Earlier observations were collected using VIGOR panorama identifiers, with additional earlier and recent street-view and satellite imagery gathered across seven named cities.Observations are paired by panorama identifier, capture time, and GPS coordinates; each period’s satellite patch is centered on its street-view capture location.Source: §3.1 · p. 4

3.2 CHANGE CLASSIFICATION AND CONSISTENCY SCREENING

• The pipeline excludes severe heading mismatches, classifies structural changes, and retains quadruplets only when every applicable cross-view and temporal consistency score reaches at least 7/10.Change classes include building change, road change, both, or no change; screening focuses on major structural changes rather than transient objects or minor appearance variations.Source: §3.2 · p. 5

3.3 LOCAL EDITING INSTRUCTIONS AND VALIDATION

• Satellite comparisons describe earlier states in four cardinal directions, while recent panoramas anchor viewpoint and unchanged appearance for local instructions.Instructions use MODIFY for a recent entity to remove and its earlier state to generate, and PRESERVE for content to retain; validation requires scores of at least 7 on Grounding, Consistency, and Preservation.Source: Figure 2 · p. 4, §3.3 · p. 5

4.1 TASK FORMULATION

• Historical street-view generation takes a recent view and a local instruction as generation inputs; the earlier view is a supervision and evaluation target, not an input.The task restores changed content while preserving content that remained unchanged.Source: §4.1 · p. 5

4.2 SUPERVISED INITIALIZATION

• Attention-projection LoRA adapters are trained on FLUX.2 [Klein] 4B with the backbone, MLPs, and encoder/decoder frozen.Changed examples use the earlier view as target; no-change examples use a PRESERVE instruction and the recent input as target, jointly training editing and preservation.Source: §4.2 · p. 6

4.3 THREE-CRITERION VLM REWARD AND EVALUATION PROTOCOL

• The shared VLM protocol scores Instruction Alignment, Background Preservation, and Quality and Physical Plausibility from 0 to 10 for both RL rewards and offline evaluation.IA checks target-referenced edit correctness; BP checks preservation outside requested edits; QP checks introduced visual defects and physical implausibility.Source: §4.3 · p. 6

4.4 MULTI-REWARD FLOW-GRPO

• GDPO normalizes each reward dimension before aggregation because direct aggregation made advantages sensitive to dimension-specific variability.Flow-GRPO samples candidates under the same condition and updates LoRA parameters using a PPO-style clipped objective; batch normalization then controls overall advantage scale, with clipping limiting extreme advantages.Source: §4.4 · pp. 6–7

5.1 EXPERIMENTAL SETUP

• The test set contains 1,094 changed samples; editors receive matched recent views and instructions, while cross-view baselines are retrained on VIGOR-his.Evaluation uses IA, BP, and QP for editing, and CV-C, CV-VR, and CV-PQ for cross-view comparisons; both use Gemini-3.1-Flash-Lite at temperature 0.Source: §5.1 · p. 7

5.2 QUANTITATIVE COMPARISONS

• On VIGOR-his, CrossTimeEdit achieves the highest IA, QP, and Overall among evaluated open-source general image editing models; versus FLUX.2 [Klein] 4B, IA and Overall improve 33.39% and 17.12%, respectively.Source: §5.2 · pp. 7–8, Table 2 · p. 8
• CrossTimeEdit substantially outperforms cross-view generation models in CV-C, CV-VR, CV-PQ, and Overall, despite lower PSNR, SSIM, and SD than Sat2Density and ControlS2S.Human preference rates correlate positively with both Editing Overall and Cross-view Overall.Source: §5.2 · p. 8, Table 3 · p. 8

5.3 QUALITATIVE COMPARISONS

• Five qualitative cases cover building, construction, and road changes, illustrating requested local edits alongside retained background and coherent, plausible outputs.The examples connect instruction following, background preservation, and image quality and physical plausibility to IA, BP, and QP.Source: §5.3 · p. 8, Figure 3 · p. 9

5.4 ABLATION STUDY

• SFT improves instruction following but reduces background preservation and image quality; RL further improves editing while restoring those qualities, and reward ablations show weaker balance without GT or dimension-wise normalization.CrossTimeEdit achieves the best balance across IA, BP, QP, and Overall.Source: §5.4 · p. 9, Table 4 · p. 9

6 CONCLUSION

• VIGOR-his contains 43,653 decade-spanning location-level quadruplets across 11 cities on three continents, and CrossTimeEdit edits recent street views using temporal satellite evidence.Training uses supervised fine-tuning and multi-reward online reinforcement learning.Source: §6 · p. 10
• Editing Overall improves 17.12% over pretrained FLUX.2 [Klein] 4B; CrossTimeEdit has the highest Overall, IA, and QP among evaluated open-source image-editing models.It also substantially outperforms cross-view generation models under the cross-view evaluation protocol.Source: §6 · p. 10
• Remaining challenges include historical details unavailable from satellite imagery, viewpoint differences across acquisition periods, and evaluator-specific preferences.The conclusion suggests richer historical evidence and independent reward and evaluation models as future directions.Source: §6 · p. 10

AI USE STATEMENT

• Generative AI assisted manuscript drafting, translation, and LaTeX preparation; the assisted content and supporting artifacts were reviewed and verified.Vision-language models also support change classification, instruction generation, screening, rewards, and evaluation.Source: AI USE STATEMENT · p. 10

• Google Street View and Google Maps imagery include face and license-plate blurring protections.Source: ETHICS STATEMENT · p. 10
• Preprocessing removes invalid and low-quality observations, and the analysis is limited to scene-level modeling rather than identifying people or inferring sensitive attributes.Source: ETHICS STATEMENT · p. 10
• The human preference study collects only aggregated comparisons, not personal or sensitive information.Source: ETHICS STATEMENT · p. 10
• Generated historical scenes are hypotheses rather than archival records and are not intended for identity, legal, or safety-critical decisions.Source: ETHICS STATEMENT · p. 10

• The paper specifies VIGOR-his construction and preprocessing, SFT and online RL objectives, and the data splits, training and inference configurations, checkpoints, reward weights, and evaluation seed.Source: REPRODUCIBILITY STATEMENT · p. 10
• Main experiments use a fixed test split and temperature-zero Gemini-3.1-Flash-Lite evaluation protocol, with manifests, configurations, prompts, and checkpoint selections stated for reproducing comparisons.Source: REPRODUCIBILITY STATEMENT · p. 10
• Implementation code, dataset metadata, and model weights are publicly available.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 SUPPLEMENTARY DETAILS FOR DATA COLLECTION AND PRE-PROCESSING

• VIGOR-source cities pair 2019 panoramas with 2008–2009 observations, while seven additional cities use 2012–2015 and 2022–2025 imagery; both branches select observations within 5 m.VIGOR-source matching keeps the shortest-distance pair for each earlier panorama; the additional-city branch clusters candidates spatially and suppresses nearby pair midpoints. Panoramas are heading-corrected, satellite crops are north-up 512 × 512 windows, and pairs are discarded if either crop is invalid.Source: §A.1 · pp. 15–16

A.2 TRAINING OBJECTIVES AND OPTIMIZATION DETAILS

• Both SFT phases share an objective sampling uniformly among T=1000 training timesteps and targeting velocity η −z⋆; online GRPO averages transition log probabilities over non-batch dimensions and applies frozen-reference velocity MSE (ε=10−3, β=2×10−3).Source: §A.2 · pp. 16–17

A.3 VISION-LANGUAGE EVALUATION PROMPTS

• Editing evaluation separates instruction alignment, background preservation, and quality/physics, while cross-view prompts separately assess scene consistency, visual realism, and perceptual quality.Editing uses a 2 × 2 input/ground-truth/output composite for IA and stacked input/output comparisons for BP and QP; scores range from 0–10. Cross-view scores range from 1–5, and both overall scores are weighted 0.45, 0.30, and 0.25 across their three criteria.Source: §A.3 · pp. 17–20
• CV-PQ rates perceptual image quality without requiring a real reference, assessing clarity, artifacts, exposure, texture, spatial consistency, panorama continuity, and viewing comfort on a 1–5 scale.The rubric ranges from severe widespread degradation (1) to consistent clarity and negligible degradation (5); it separates semantic or physical implausibility unless it also causes a visible rendering defect.Source: §A.3 · pp. 23–24

B.1 COMPLETE TRAINING CONFIGURATION

• The SFT manifest contains 8,797 changed and 13,196 no-change samples; RL samples about 2,000 changed conditions and initializes from the epoch-9 SFT checkpoint, also used as its frozen reference.SFT uses about 13,750 optimizer updates across two phases; RL and evaluation specify seed 42, while the SFT launch scripts do not fix a seed.Source: §B.1 · p. 24

B.2 IMAGE SIMILARITY METRICS

• Generated panoramas are compared with location-matched earlier ground truth across the same 1,094 test locations, using PSNR, SSIM, LPIPS, and Sharpness Difference.PSNR and SSIM are higher-is-better, LPIPS is lower-is-better, and Sharpness Difference measures correspondence in local edge-strength patterns, with higher scores indicating closer patterns.Source: §B.2 · pp. 25–26
• References, not generated images, are bicubic-resized from 2048 × 1024 to 1024 × 512; SSIM uses RGB channels, data_range=1.0, and a 7 × 7 uniform window, while LPIPS uses version-0.1 weights and AlexNet.The separately computed Gaussian-window SSIM diagnostic is excluded from reported results.Source: §B.2 · p. 25

B.3 MAIN EXPERIMENT: REWARD AND VALIDATION CURVES

• IA, BP, and QP reward trajectories rose overall from RL epoch 1 to 240 despite noticeable fluctuations.Figure 4 reports mean rewards over valid candidates and mean scores on the full 1,095-location change validation split.Source: §B.3 · p. 26, Figure 4 · p. 26
• Epoch 200 attained the highest weighted Overall among evaluated RL checkpoints (7.901), with IA 7.856, BP 8.255, and QP 7.559.Epoch 240 had slightly higher IA (7.873), but lower BP (7.739), QP (7.436), and weighted Overall (7.723); epoch 200 was selected as the final model. Each checkpoint had 1,095 valid scores per dimension, with no failed requests.Source: Figure 4 · p. 26, §B.3 · p. 26

B.4 HUMAN PREFERENCE STUDY

• Fifteen volunteers with geography-related academic backgrounds judged randomized model pairs; majority votes yielded 50 comparisons per model and model-level human win rates.The study evaluated 11 models through all 55 unordered pairs, with five comparisons per pair.Source: §B.4 · p. 26
• 64%, 30%, 14%, and 0% human win rates correspond to CrossTimeEdit, ControlS2S, Sat2Density, and ControlNet; their Cross-view Overall scores are 3.546, 2.084, 2.015, and 1.740.For these four models, the VLM scores also yielded Pearson r = 0.896 and Spearman ρ = 1.000 against human preferences.Source: Table 7 · p. 27, §B.4 · p. 27
• Editing Overall correlated positively with human win rates (Pearson r = 0.804; Spearman ρ = 0.843), with 22 of 28 model pairs concordant.CrossTimeEdit ranked fourth in both human preference and Editing Overall; Qwen-Image-3.0-Pro led human preference, while GPT-Image-2 led Editing Overall.Source: §B.4 · p. 27

B.5 BASELINE CONFIGURATIONS AND COMPARISON PROTOCOLS

• General image-editing models used official weights and inference configurations with the same recent street views and local editing instructions.Cross-view models instead took earlier satellite images as input; Sat2Density, ControlS2S, and ControlNet were retrained on VIGOR-his.Source: §B.5 · p. 28
• Sat2Density used epoch 25, ControlS2S used step 4,500, and ControlNet used its final epoch-50 checkpoint for test inference.Sat2Density and ControlS2S checkpoints were selected using validation scores; ControlNet used the final checkpoint.Source: §B.5 · p. 28

C.1 VIGOR-HIS DATASET STRUCTURE AND EXAMPLES

• VIGOR-his examples pair earlier and recent satellite images with co-located street-view panoramas, then assign panorama regions MODIFY or PRESERVE operations.The examples span no change, building changes, road changes, and combined building and road changes; instructions organize regions by viewing direction.Source: §C.1 · p. 29, Figure 6 · p. 29

C.2 ABLATION STUDY: QUALITATIVE COMPARISON

• SFT-only (w/o RL) raises IA to 6.919 but lowers BP/QP to 6.698/6.543, while Baseline and w/o SFT preserve backgrounds better but align poorly.Baseline and w/o SFT have BP scores of 8.008 and 8.112, and IA scores of 5.960 and 6.085, respectively; Fig. 7 reflects the same trade-off.Source: §C.2 · pp. 29–30, Figure 7 · p. 30
• Full CrossTimeEdit reaches IA 7.950, BP 8.264, and QP 7.631; w/o GT and w/o DGN approach its IA but trail in BP/QP and background reliability.Qualitative cases show Full preserving non-target road and building surfaces while restoring or removing the requested structures.Source: Figure 7 · p. 30, §C.2 · p. 30

C.3 QUALITATIVE COMPARISON WITH CROSS-VIEW GENERATION MODELS

• CrossTimeEdit edits from satellite-derived change instructions while using recent street views to constrain viewpoint and unchanged appearance, producing outputs closer to earlier street-view observations than cross-view generation models.Fig. 8 shows the cross-view models differing from ground truth in viewpoint, scene structure, and appearance.Source: §C.3 · p. 30, Figure 8 · p. 31

D DATASET-CONSTRUCTION PROMPTS

• Cross-view and temporal audits check structural consistency across street-view and satellite imagery; all applicable consistency scores must be at least 7, and blurred street views receive a fatal zero score.The prompts evaluate permanent structures and land cover while disregarding temporary objects and visibility-inaccessible rooftop details.Source: D DATASET-CONSTRUCTION PROMPTS · pp. 31, 33, 34
• Dataset prompts screen out major viewpoint misalignment, ignore transient or minor differences, and classify clear building and road changes as core targets.The prompts specify examples such as building demolition or major renovation, new crosswalks, conspicuous pavement updates, and road-structure changes.Source: D DATASET-CONSTRUCTION PROMPTS · pp. 31–32
• Satellite comparisons report permanent structural changes visible from street level in four directions, while ignoring weather, seasonal, vehicle, and image-quality differences.The prompts cover buildings and road changes, require detailed descriptions of pre-change structures, and exclude changes visible only from above.Source: D DATASET-CONSTRUCTION PROMPTS · p. 38
• Local editing instructions map satellite-confirmed changes to panorama regions and direct the model to modify changed areas or preserve unchanged ones, accounting for occlusion and perspective.The panorama maps north to the center, east and west to the right and left quarters, and south to both side edges; occluded changes are marked PRESERVE.Source: D DATASET-CONSTRUCTION PROMPTS · pp. 39–40
• Consistency checks score whether MODIFY targets are visible at the stated location in the current view and whether restored entities match the historical view, with each score governed by its worst block.Executability and historical consistency are each scored from 0 to 10; the former checks the current panorama and the latter checks the historical target.Source: D DATASET-CONSTRUCTION PROMPTS · p. 41
• Preservation checks whether regions labeled PRESERVE remain structurally unchanged between views, treating permanent tree changes as structural but seasonal changes as incidental.Any changed PRESERVE region lowers the overall score; if every region is marked MODIFY, the preservation score is 10.Source: D DATASET-CONSTRUCTION PROMPTS · p. 42

E FAILURE-CASE ANALYSIS AND LIMITATIONS

• Figure 9 illustrates these limits: a crosswalk is not clearly visible from the street viewpoint, and a complex building is ambiguous from satellite evidence.In the top case, the earlier satellite image shows a construction site while the street-view ground truth already shows the completed building.Source: E FAILURE-CASE ANALYSIS AND LIMITATIONS · p. 43, Figure 9 · p. 43
• Cross-view visibility gaps, residual temporal-state mismatches, and insufficient facade evidence constrain reliable historical street-view recovery.Occlusion can hide satellite-visible features at street level, while complex buildings may yield underspecified instructions and multiple plausible outputs.Source: E FAILURE-CASE ANALYSIS AND LIMITATIONS · p. 43

F FUTURE WORK

• Future work could add historical maps, aerial imagery, building footprints, archival photographs, and ground-level records to inform satellite-invisible regions.The current framework derives temporal change evidence primarily from paired satellite observations, which do not fully reveal facades, storefronts, road-level details, or occluded structures.Source: F FUTURE WORK · p. 43
• Future studies could use independent reward models, multiple-VLM ensembles, or human-calibrated evaluators to reduce evaluator-specific preferences and over-optimization toward one protocol.CrossTimeEdit currently uses the same VLM protocol for reinforcement-learning rewards and evaluation.Source: F FUTURE WORK · p. 43
• The current formulation generates earlier views from recent views and satellite evidence; future work could reverse time direction or extend to continuous, multi-temporal observations.The proposed extensions target present-day generation from historical views and modeling longer-term urban evolution beyond a single earlier–recent transition.Source: F FUTURE WORK · p. 43