acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Historical street-view imagery records urban evolution, but uneven coverage leaves substantial gaps in historical records. Generating plausible past appearances requires restoring changed structures while preserving persistent scene content. We construct VIGOR-his, a decade-spanning cross-view dataset containing 43,653 location-level quadruplets across 11 cities on three continents. Its automated pipeline performs spatial pairing, consistency screening, change classification, and the generation and validation of satellite-based change descriptions and local editing instructions. Based on VIGOR-his, we propose CrossTimeEdit, a model that reformulates historical street-view generation as editing, using recent street views to constrain viewpoint and unchanged appearance and temporal satellite differences as change evidence. Starting from FLUX.2 [Klein] 4B, we train CrossTimeEdit through supervised fine-tuning (SFT) followed by online reinforcement learning (RL). We design three street-view editing criteria, namely Instruction Alignment (IA), Background Preservation (BP), and Quality and Physical Plausibility (QP), as both RL reward dimensions and evaluation metrics. We optimize this multi-reward objective using Within Group Relative Policy Optimization for flow-matching models (Flow-GRPO) with Group reward-Decoupled Normalization Policy Optimization (GDPO), which normalizes each reward dimension before aggregation. CrossTimeEdit improves overall performance across the three editing criteria by 17.12% over the pretrained baseline and outperforms cross-view generation models in scene consistency, visual realism, and perceptual quality. The implementation code, dataset, and model weights are available at https://anonymous.4open.science/r/CrossTimeEdit2027/.

Primary area:
generative models
Keywords:
Satellite Imagery, Street-View Imagery, Generative Models, Datasets, Vision-Language Models, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_d2xQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces VIGOR-his, containing 43,653 temporal street-view and satellite quadruplets across 11 cities. CrossTimeEdit uses recent street views and satellite-derived instructions to generate historical appearances through supervised fine-tuning and multi-reward reinforcement learning. The value of a dataset may be greater than the work presented in the paper itself.

Critical strengths
  1. This dataset fills a gap in historical cross-view observation.
  2. The authors' editing scheme accounts for the use of historical information and the preservation of context, employing a reward mechanism to evaluate the execution of edits, content retention, and editing quality.
  3. The ablation studies are comprehensive.
Critical weaknesses
  1. In the context of buildings, human activity accounts for a significant portion and inherently involves a degree of randomness. Consequently, the images generated by the model may appear plausible while failing to align with reality—as illustrated by the results in Figure 3.

  2. It appears that the model used to provide training rewards is the same as the one used for final evaluation. This raises concerns as to whether the reinforcement learning (RL) process is being employed solely to boost specific metrics.

  3. Regarding the baseline methods, some of the training-based approaches cited are somewhat dated—for instance, ControlNet based on SD1.5. It would be more appropriate to include comparisons based on the same foundational model.

Questions for authors

Do existing human judgments or independent evaluators confirm the SFT-to-RL improvement on requested historical changes?

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

Visual plausibility and historical reality (Weakness 1)

We agree with the reviewer that human activity introduces variability that makes the exact recovery of historical street scenes challenging, particularly for transient objects such as pedestrians and vehicles. Historical satellite imagery is limited by spatial resolution, overhead viewing geometry, and occlusion, and its acquisition time generally differs from that of the corresponding street view. Consequently, it cannot reliably establish which pedestrians or vehicles were present at a particular street-level moment, their locations, or their appearance. Without sufficient observational evidence, a VLM cannot reliably describe these transient states or translate them into historically grounded editing instructions and learning signals.

CrossTimeEdit therefore focuses on persistent building and road changes that can be supported by temporal satellite evidence. Its annotation pipeline disregards transient objects when identifying structural changes, while the recent street view provides a reference for viewpoint and persistent scene content. This formulation directs learning toward restoring the historical semantic layout of buildings and roads, rather than reconstructing the precise configuration of human activity at a particular moment.

Despite the uncertainty in transient human activity, the qualitative results in Fig. 3 show that CrossTimeEdit can recover the building and road configurations that characterize the earlier scene. The relevant evidence is the restoration of changed structures and the retention of surrounding scene content, rather than the presence of particular pedestrians or vehicles. Compared with the direct cross-view generation methods illustrated in Appendix C.3, using the recent street view as a reference helps retain observed spatial context and concentrate generation on the historical changes supported by satellite evidence. This enables CrossTimeEdit to recover a semantic layout more consistent with the earlier street view, even when moment-specific human activity cannot be reconstructed.

We have clarified in Section 4.1 that the task targets persistent building and road changes rather than the exact reconstruction of transient objects such as pedestrians and vehicles. Section 5.3 now emphasizes that the qualitative results demonstrate visually plausible generation together with recovery of the historical semantic layout. We have also expanded Appendix E to discuss the observational constraints on historical reconstruction, and Appendix F to consider archival photographs, temporally and geographically grounded textual records, and uncertainty modeling as directions for incorporating additional historical evidence and representing plausible alternatives.

AuthorsOpenReview ↗

Response to Weakness 2 and Question

Independent confirmation of SFT-to-RL improvements (Weakness 2 and Question)

We thank the reviewer for requesting human or independent evaluation of the improvements from SFT to RL. We agree that verification beyond the training reward model is important for establishing whether RL improves historical editing performance. We therefore conducted evaluations using GPT-5.6-Terra and Gemini-3.6-Flash, both separate from the Gemini-3.1-Flash-Lite reward model, and provide complementary qualitative ablation comparisons.

The table below reports the independent ablation results, including a direct comparison between SFT alone (w/o RL) and the full SFT-plus-RL model. Both evaluators assess the same generated outputs using the original IA/BP/QP criteria and Overall weights of (0.45, 0.30, 0.25).

GPT-5.6-Terra

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline5.8278.4677.3186.992
w/o RL6.2926.4877.3616.618
w/o SFT6.0388.5427.3837.126
w/o GT6.6656.6907.4336.865
w/o DGN6.7506.8977.3926.954
Full6.8937.4937.5487.237

Gemini-3.6-Flash

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline4.5616.5036.4995.628
w/o RL5.0535.2065.9975.335
w/o SFT4.6136.3886.4105.595
w/o GT4.7885.1546.3825.296
w/o DGN4.9585.3106.4485.436
Full5.8385.9206.6476.065

The results show consistent improvements from SFT to RL across all three criteria under both independent evaluators. Under GPT-5.6-Terra, IA increases from 6.292 to 6.893, BP from 6.487 to 7.493, QP from 7.361 to 7.548, and Overall from 6.618 to 7.237. Under Gemini-3.6-Flash, the corresponding improvements are 5.053 to 5.838 for IA, 5.206 to 5.920 for BP, 5.997 to 6.647 for QP, and 5.335 to 6.065 for Overall. Both evaluators also rank the full model highest in Overall among all ablation variants. These results demonstrate that RL improves historical edit execution, background preservation, and visual quality beyond the preferences of the training reward model.

The qualitative ablation comparisons provide complementary visual evidence. Across matched locations, the full model executes the requested structural edits while better preserving surrounding road and building appearance than the SFT-only and other ablated variants. Together, the independent scores and qualitative comparisons support consistent improvements in historical editing from SFT to RL.

The table below further compares CrossTimeEdit with general image editing models on VIGOR-his. CrossTimeEdit achieves the highest IA, QP, and Overall among the evaluated open-source image editing models under both independent evaluators, showing that its competitive performance also persists beyond the original evaluator.

GPT-5.6-Terra

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite8.3179.0667.6098.365
Qwen-Image-3.0-Pro7.6299.4087.6588.170
GPT-Image-27.9499.3777.8048.341
Open-source models
FLUX.2 [Klein] 4B5.8278.4677.3186.992
FLUX.2 [Klein] 9B6.4807.8397.2407.078
OmniGen25.3208.1156.5426.464
LongCat-Image-Edit6.5698.0277.1277.146
CrossTimeEdit6.8937.4937.5487.237

Gemini-3.6-Flash

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite5.6996.6016.7386.229
Qwen-Image-3.0-Pro5.7796.9576.8566.401
GPT-Image-26.1787.1097.0466.674
Open-source models
FLUX.2 [Klein] 4B4.5616.5036.4995.628
FLUX.2 [Klein] 9B4.8695.9396.3365.557
OmniGen23.7206.1155.9414.994
LongCat-Image-Edit4.9926.1546.4515.705
CrossTimeEdit5.8385.9206.6476.065

We have added the independent evaluation protocol and analysis to Section 5.5, with the main results summarized in Fig. 4. The complete model comparison and ablation results are reported in Tables 5 and 6 of Appendix B.2, respectively. The qualitative ablation comparisons and accompanying analysis are provided in Appendix C.2.

AuthorsOpenReview ↗

Response to Weakness 3

Baseline selection and comparison with the same backbone (Weakness 3)

We thank the reviewer for this suggestion. Recent cross-view methods such as CrossViewDiff [1] and SatDreamer360 [2] build on Stable Diffusion-based architectures. To include a stronger backbone in the comparison, we fine-tuned FLUX.2 [Klein] 4B with LoRA on VIGOR-his for satellite-to-street generation, using the earlier satellite image as input and the same pretrained foundation model as CrossTimeEdit.

Starting from the pretrained distilled model, we trained Transformer LoRA adapters with rank and alpha both set to 32, while freezing the backbone weights, VAE, and text encoder. Training used AdamW with a learning rate of 3 × 10⁻⁵ and a global effective batch size of 8 on four NVIDIA RTX 3090 GPUs for 10 epochs. The final checkpoint generated 1024 × 512 panoramas using a fixed satellite-to-street prompt, four denoising steps, and a guidance scale of 1.0.

The table below reports the updated comparison with cross-view generation models on VIGOR-his.

ModelPSNR ↑SSIM ↑LPIPS ↓SD ↑CV-C ↑CV-VR ↑CV-PQ ↑Overall ↑
Sat2Density14.6850.5320.63420.0402.0481.9732.0062.015
ControlS2S14.1190.5190.65720.0152.0272.1802.0712.084
ControlNet12.1420.3440.68017.8851.1192.1182.4041.740
FLUX.2 [Klein] 4B14.7160.5260.59219.3252.1601.4951.8871.892
CrossTimeEdit12.5890.4630.53018.7053.0053.9803.9983.546

CrossTimeEdit outperforms the fine-tuned FLUX.2 [Klein] 4B baseline in all three cross-view evaluation criteria, achieving CV-C of 3.005 versus 2.160, CV-VR of 3.980 versus 1.495, and CV-PQ of 3.998 versus 1.887. Overall increases from 1.892 to 3.546, while LPIPS decreases from 0.592 to 0.530. These results show that CrossTimeEdit’s advantages persist when the pretrained foundation model is shared, supporting the effectiveness of using recent street views to preserve persistent scene content while applying historical edits guided by temporal satellite evidence.

We have added this baseline to Table 3, introduced its cross-view fine-tuning setup in Section 5.1, and updated the quantitative comparison in Section 5.2. Appendix B.8 now documents its training and inference configurations.

[1] Weijia Li et al. “CrossViewDiff: A Cross-View Diffusion Model for Satellite-to-Street View Synthesis.” Pattern Recognition, 114735, 2026.

[2] Xianghui Ze et al. “SatDreamer360: Geometry Consistent Street-View Video Generation from Satellite Imagery.” arXiv preprint, 2025.

Reviewer_vm1qOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CrossTimeEdit, a framework for historical street-view generation, together with VIGOR-his, a decade-spanning temporal cross-view dataset. The dataset contains 43,653 location-level quadruplets of historical and recent satellite and street-view images across 11 cities. The authors construct the dataset through spatial and temporal matching, VLM-based consistency screening, change classification, and automatic editing instruction generation.

Rather than generating historical street views directly from satellite images, CrossTimeEdit formulates the task as image editing. It uses recent street views to preserve unchanged scene content and satellite-derived instructions to restore historical structures. Starting from FLUX.2 Klein 4B, the model is trained through supervised fine-tuning followed by multi-reward online reinforcement learning using Flow-GRPO and dimension-wise reward normalization.

Experiments against general image editing and cross-view generation methods demonstrate improvements in instruction alignment, perceptual quality, and scene consistency. Ablation studies and a human preference study provide additional evidence for the proposed training strategy.

Critical strengths
  • The paper addresses a meaningful problem: recovering plausible historical street-view appearances from contemporary observations and temporal satellite evidence. Formulating this problem as image editing rather than generation from scratch is intuitive and well motivated.
  • The construction of VIGOR-his is also a valuable contribution. Collecting and aligning temporal street-view and satellite observations across 11 cities involves nontrivial data processing. The resulting dataset could support further research on temporal cross-view understanding, urban change analysis, and historical scene generation.
  • CrossTimeEdit demonstrates improvements over the pretrained backbone and several open-source image editing models. The ablation studies also provide useful insights into the training strategy. SFT improves instruction following but weakens background preservation, while subsequent RL improves the balance among instruction alignment, preservation, and visual quality. The improvements in image similarity metrics and the positive results from the human preference study provide additional support beyond the VLM evaluation alone.
Critical weaknesses
  • My main concern is that the paper uses the same Gemini-3.1-Flash-Lite model and IA/BP/QP evaluation protocol for both RL rewards and final editing evaluation. Consequently, improvements in these metrics may partly reflect optimization toward the preferences of the specific reward model, rather than improvements in historical reconstruction fidelity. The authors attempt to address this concern through a human preference study, which is appreciated. However, the study involves only five image comparisons per model pair, and the reported correlations primarily establish agreement in model-level preferences. They do not necessarily demonstrate that the generated historical structures accurately match the actual past state of each location.
  • While the scale of VIGOR-his is impressive, the dataset construction relies heavily on VLM-based screening, change classification, instruction generation, and instruction validation. The paper does not provide a sufficiently comprehensive independent human audit of the resulting annotations. It would be helpful to quantify the accuracy of temporal correspondence, change categories, and generated editing instructions on a manually verified subset. This is important because annotation errors or incomplete change descriptions could directly affect both training and evaluation. Additionally, the paper reports 43,653 collected quadruplets, whereas the main experiments use 21,993 training samples and approximately 1,100 validation and test samples each. The relationship between the complete dataset and these experimental subsets should be clarified.
Questions for authors
  • Since the same VLM and scoring protocol are used for RL rewards and final editing evaluation, could the authors report results using an independent evaluator? In particular, could historical structural correctness be evaluated separately from general visual preference?
  • Have the authors conducted a manually verified quality assessment of VIGOR-his? What are the accuracy rates of temporal pairing, change classification, and editing instruction generation? How many candidate samples were discarded at each VLM screening stage?
  • How do the 43,653 collected quadruplets relate to the approximately 24,000 samples used in the reported training, validation, and testing configurations?
6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Independent evaluation and historical correctness (Weakness 1 and Question 1) — Part 1/2

We thank the reviewer for highlighting the distinction between adapting to an evaluator’s preferences and improving historical street-view reconstruction. We agree that using Gemini-3.1-Flash-Lite for both training rewards and evaluation warrants verification with independent evaluators. Our human preference study establishes agreement between human judgments and Gemini-3.1-Flash-Lite at the level of model rankings, but does not directly establish correspondence with the actual historical scene. Independent evaluation using historical references therefore provides complementary evidence for assessing whether the improvements extend beyond the training reward model’s visual preferences.

We therefore conducted independent evaluations using GPT-5.6-Terra and Gemini-3.6-Flash, covering comparisons with general image editing models and the ablation variants. Both evaluators assess the same generated outputs using the original evaluation inputs, IA/BP/QP criteria, and Overall weights of (0.45, 0.30, 0.25).

The table below compares CrossTimeEdit with general image editing models on VIGOR-his under both independent evaluators.

GPT-5.6-Terra

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite8.3179.0667.6098.365
Qwen-Image-3.0-Pro7.6299.4087.6588.170
GPT-Image-27.9499.3777.8048.341
Open-source models
FLUX.2 [Klein] 4B5.8278.4677.3186.992
FLUX.2 [Klein] 9B6.4807.8397.2407.078
OmniGen25.3208.1156.5426.464
LongCat-Image-Edit6.5698.0277.1277.146
CrossTimeEdit6.8937.4937.5487.237

Gemini-3.6-Flash

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite5.6996.6016.7386.229
Qwen-Image-3.0-Pro5.7796.9576.8566.401
GPT-Image-26.1787.1097.0466.674
Open-source models
FLUX.2 [Klein] 4B4.5616.5036.4995.628
FLUX.2 [Klein] 9B4.8695.9396.3365.557
OmniGen23.7206.1155.9414.994
LongCat-Image-Edit4.9926.1546.4515.705
CrossTimeEdit5.8385.9206.6476.065

The model comparison shows that CrossTimeEdit achieves the highest IA, QP, and Overall among the evaluated open-source image editing models under both independent evaluators. Relative to the pretrained FLUX.2 [Klein] 4B baseline, IA increases from 5.827 to 6.893 under GPT-5.6-Terra and from 4.561 to 5.838 under Gemini-3.6-Flash, while QP increases from 7.318 to 7.548 and from 6.499 to 6.647, respectively. Overall improves from 6.992 to 7.237 and from 5.628 to 6.065, corresponding to gains of 3.50% and 7.76%. These results jointly support improved execution of historical edits, visual quality, physical plausibility, and overall editing performance under independent assessment. BP exhibits evaluator-dependent behavior, reflecting differences in the assessment of background preservation alongside these consistent gains.

AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Independent evaluation and historical correctness (Weakness 1 and Question 1) — Part 2/2

The following tables report the complete ablation results under the same independent evaluators used in Part 1.

GPT-5.6-Terra

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline5.8278.4677.3186.992
w/o RL6.2926.4877.3616.618
w/o SFT6.0388.5427.3837.126
w/o GT6.6656.6907.4336.865
w/o DGN6.7506.8977.3926.954
Full6.8937.4937.5487.237

Gemini-3.6-Flash

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline4.5616.5036.4995.628
w/o RL5.0535.2065.9975.335
w/o SFT4.6136.3886.4105.595
w/o GT4.7885.1546.3825.296
w/o DGN4.9585.3106.4485.436
Full5.8385.9206.6476.065

The ablation results further show that the full model achieves the highest Overall among all variants under both independent evaluators. Adding RL after SFT improves IA, BP, and QP compared with the SFT-only variant (w/o RL), with Overall increasing from 6.618 to 7.237 and from 5.335 to 6.065, respectively. These consistent improvements confirm that the benefits of the training strategy also carry over to independent evaluation.

CrossTimeEdit’s improvements persist across independent evaluators, extending beyond agreement with Gemini-3.1-Flash-Lite’s visual preferences. The consistent improvements across complementary criteria, evaluated with historical references, provide strong supporting evidence for improved historical structural correctness and robust editing performance. We have added the independent evaluation protocol and analysis to Section 5.5, summarized the main results in Fig. 4, and provided the complete model comparison and ablation results in Tables 5 and 6 of Appendix B.2.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2

Human audit and screening statistics (Weakness 2 and Question 2)

We thank the reviewer for requesting human verification of the annotations and a clearer account of the samples discarded during dataset construction. We agree that these complementary assessments are important for establishing annotation reliability and making the screening process transparent. We address both aspects below.

  1. Independent human audit of the annotations. We conducted a human audit of 200 randomly sampled quadruplets, stratified equally across building change, road change, building-and-road change, and no change, and covering all 11 cities. Four participants with backgrounds in geography and urban studies assessed spatiotemporal pairing, within-period state consistency, change categories, and instruction grounding. The audit explicitly separates location correspondence and viewpoint comparability from agreement of the building, road, and development states between street and satellite views at each acquisition period.

    The table below shows 93.0% overall agreement, with agreement of 88.0–100.0% across the four categories. This consistently high agreement supports the reliability of the change annotations across both changed and unchanged scenes.

    CategoryAgreed / auditedAgreement
    Building change45 / 5090.0%
    Road change44 / 5088.0%
    Building-and-road change50 / 50100.0%
    No change47 / 5094.0%
    Overall186 / 20093.0%

    The table below reports the graded quality assessments, where a score of 3 denotes full consistency. All three dimensions achieve mean scores of at least 2.800 out of 3, and at least 98.7% of samples receive scores of 2 or above, indicating at least substantial consistency. Spatiotemporal pairing and within-period state consistency are assessed on all 200 quadruplets, while instruction grounding is assessed on the 150 changed samples.

    DimensionnMean (0–3)Score 3Score ≥2
    Spatiotemporal pairing2002.86587.5%99.0%
    Within-period state consistency2002.90591.0%99.5%
    Instruction grounding1502.80081.3%98.7%

    Together with the category-agreement results, these findings demonstrate the high quality of VIGOR-his: its change annotations are reliable, its spatiotemporal pairings are accurate, its within-period scene states are consistent, and its editing instructions are well grounded in satellite evidence. These properties make VIGOR-his suitable for training and evaluating historical street-view editing models.

  2. Candidates discarded at each screening stage. Viewpoint Screening and Change Classification (Section 3.2) removes 1,014 of 76,039 candidate quadruplets because of severe heading misalignment (category E), retaining 19,810 changed samples (categories A, B, and C) and 55,215 no-change samples (category F).

    Cross-View and Temporal Consistency Screening (Section 3.2) removes 4,666 changed samples and 22,548 no-change samples, retaining 15,144 and 32,667 samples, respectively. Instruction Consistency Validation (Section 3.3) then removes 4,158 changed samples, retaining 10,986. The resulting dataset therefore contains 10,986 changed and 32,667 no-change quadruplets, totaling 43,653.

We have added the human-audit protocol and results to Appendix B.1, with its purpose and main conclusion summarized in Section 3. The numbers of samples removed and retained at each screening stage are now reported at the beginning of Appendix D, using the same step names as Section 3.

Reviewer_6znuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CrossTimeEdit, a framework for historical street-view generation using temporal satellite imagery and recent street-view observations. The authors introduce VIGOR-his, a decade-spanning cross-view dataset containing 43,653 location-level quadruplets across 11 cities, with satellite-grounded change descriptions and local editing instructions. Instead of generating historical street views from scratch, the proposed method formulates the task as conditional image editing, using recent street views to preserve unchanged scene content while restoring historical structures. The model is trained through supervised fine-tuning followed by multi-reward reinforcement learning, optimizing instruction alignment, background preservation, and visual quality. Experiments demonstrate improvements over existing open-source image editing models and cross-view generation baselines.

Critical strengths
  • The VIGOR-his dataset

The construction of VIGOR-his addresses an important gap in existing cross-view datasets by providing temporally aligned street-view and satellite imagery spanning approximately a decade. Its geographic coverage, structured change annotations, and local editing instructions make it a potentially valuable resource for future research in historical scene generation and urban change understanding.

  • Well-motivated framework with comprehensive experimental validation

Reformulating historical street-view generation as localized image editing is a sensible approach that helps preserve viewpoint and unchanged scene content. The combination of supervised fine-tuning and multi-reward reinforcement learning effectively balances instruction following, background preservation, and visual quality. Extensive comparisons with existing methods, supported by ablation studies and human preference evaluations, demonstrate the practical effectiveness of the proposed framework.

Critical weaknesses
  • Lack of human validation for dataset quality

The VIGOR-his dataset is constructed largely through automated VLM-based screening, change classification, instruction generation, and consistency validation. However, the paper does not provide a systematic human evaluation of the dataset's annotation accuracy and overall quality. Given the extensive reliance on VLM-generated annotations and judgments, human verification of a representative subset would be important to establish the reliability of the dataset.

  • Insufficient analysis of multi-reward RL optimization

The proposed RL framework combines Flow-GRPO with dimension-wise reward normalization and empirically selected weights for instruction alignment, background preservation, and visual quality. Although the ablation study demonstrates the benefit of DGN, the paper does not investigate alternative reward aggregation strategies or the sensitivity to reward weights. Furthermore, Table 4 shows that removing DGN improves instruction alignment while reducing the other metrics, suggesting a trade-off among objectives. Additional comparisons would help clarify the effectiveness and robustness of the proposed multi-reward optimization.

  • Ambiguity in reconstructing historical appearance from satellite evidence

Satellite imagery provides useful information about structural changes but lacks detailed observations of building facades, materials, and occluded objects. Consequently, multiple historical street-level appearances may be equally consistent with the available satellite observations, making accurate reconstruction inherently underconstrained. To address this limitation, the authors could incorporate additional textual descriptions or archival photographs to provide richer historical appearance constraints. Alternatively, modeling uncertainty or generating multiple plausible reconstructions could better capture the ambiguity inherent in this task.

Questions for authors

See above, to summarize:

  1. Dataset quality: Has the quality of VIGOR-his been independently validated by human annotators? If so, what are the annotation accuracy and consistency across different change categories?

  2. VLM evaluation bias: How robust are the reported improvements when evaluated using an independent VLM rather than the same model used for RL rewards?

  3. RL reward design: How sensitive is the proposed multi-reward optimization to the empirically selected reward weights?

7 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Human validation of VIGOR-his (Weakness 1 and Question 1)

We thank the reviewers for emphasizing the importance of validating the quality of an automatically constructed dataset through human assessment. We agree that human verification is necessary to assess both the reliability of the change categories and the consistency of the cross-view annotations and editing instructions.

We conducted a human audit by randomly sampling 200 location-level quadruplets without replacement, with 50 from each of the four original categories and coverage of all 11 cities. Four participants with backgrounds in geography and urban studies assessed spatiotemporal pairing, within-period state consistency, change categories, and instruction grounding. Category judgments used only the two street views, while instruction grounding was assessed from the two satellite views and the editing instruction.

The audit addresses two complementary aspects:

  1. Reliability of the change categories. The table below shows 93.0% overall agreement, with agreement of 88.0–100.0% across the four categories. This consistently high agreement supports the reliability of the change annotations across both changed and unchanged scenes.

    CategoryAgreed / auditedAgreement
    Building change45 / 5090.0%
    Road change44 / 5088.0%
    Building-and-road change50 / 50100.0%
    No change47 / 5094.0%
    Overall186 / 20093.0%
  2. Consistency of cross-view annotations and editing instructions. The table below reports the graded quality assessments, where a score of 3 denotes full consistency. All three dimensions achieve mean scores of at least 2.800 out of 3, and at least 98.7% of samples receive scores of 2 or above, indicating at least substantial consistency. Spatiotemporal pairing and within-period state consistency are assessed on all 200 quadruplets, while instruction grounding is assessed on the 150 changed samples.

    DimensionnMean (0–3)Score 3Score ≥2
    Spatiotemporal pairing2002.86587.5%99.0%
    Within-period state consistency2002.90591.0%99.5%
    Instruction grounding1502.80081.3%98.7%

Together with the category-agreement results, these findings demonstrate the high quality of VIGOR-his: its change annotations are reliable, its spatiotemporal pairings are accurate, its within-period scene states are consistent, and its editing instructions are well grounded in satellite evidence. These properties make VIGOR-his suitable for training and evaluating historical street-view editing models.

We have added the complete human-audit protocol and results to Appendix B.1, including the sampling procedure, participant backgrounds, evaluation inputs, category-agreement results, and graded quality assessments. The purpose and main conclusion of the audit are also summarized in Section 3.

AuthorsOpenReview ↗

Response to Question 2

Robustness under independent evaluators (Question 2)

We thank the reviewer for suggesting independent evaluation. We agree that evaluating the model with judges other than the training reward model is important for determining whether the observed gains generalize beyond that model’s preferences. We therefore evaluated the generated outputs using GPT-5.6-Terra and Gemini-3.6-Flash. Both evaluators use the same IA/BP/QP criteria and Overall weights (0.45, 0.30, 0.25), allowing the evaluator to change while retaining the evaluation definition.

The table below shows that CrossTimeEdit achieves the highest IA, QP, and Overall among the evaluated open-source image editing models under both evaluators. Its Overall is 7.237 versus 6.992 for the pretrained FLUX.2 [Klein] 4B baseline under GPT-5.6-Terra, and 6.065 versus 5.628 under Gemini-3.6-Flash, corresponding to improvements of 3.50% and 7.76%.

GPT-5.6-Terra

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite8.3179.0667.6098.365
Qwen-Image-3.0-Pro7.6299.4087.6588.170
GPT-Image-27.9499.3777.8048.341
Open-source models
FLUX.2 [Klein] 4B5.8278.4677.3186.992
FLUX.2 [Klein] 9B6.4807.8397.2407.078
OmniGen25.3208.1156.5426.464
LongCat-Image-Edit6.5698.0277.1277.146
CrossTimeEdit6.8937.4937.5487.237

Gemini-3.6-Flash

ModelIA ↑BP ↑QP ↑Overall ↑
Closed-source models
Nano Banana 2 Lite5.6996.6016.7386.229
Qwen-Image-3.0-Pro5.7796.9576.8566.401
GPT-Image-26.1787.1097.0466.674
Open-source models
FLUX.2 [Klein] 4B4.5616.5036.4995.628
FLUX.2 [Klein] 9B4.8695.9396.3365.557
OmniGen23.7206.1155.9414.994
LongCat-Image-Edit4.9926.1546.4515.705
CrossTimeEdit5.8385.9206.6476.065

The table below reports the complete ablation results. The full model achieves the highest Overall among all ablation variants under both independent evaluators. Under both evaluators, adding RL after SFT improves IA, BP, and QP compared with the SFT-only variant (w/o RL). These consistent gains across evaluators demonstrate that the benefits of our training strategy generalize beyond the reward model’s preferences, supporting the effectiveness of the framework’s components and the robustness of its overall performance.

GPT-5.6-Terra

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline5.8278.4677.3186.992
w/o RL6.2926.4877.3616.618
w/o SFT6.0388.5427.3837.126
w/o GT6.6656.6907.4336.865
w/o DGN6.7506.8977.3926.954
Full6.8937.4937.5487.237

Gemini-3.6-Flash

MethodIA ↑BP ↑QP ↑Overall ↑
Baseline4.5616.5036.4995.628
w/o RL5.0535.2065.9975.335
w/o SFT4.6136.3886.4105.595
w/o GT4.7885.1546.3825.296
w/o DGN4.9585.3106.4485.436
Full5.8385.9206.6476.065

We have added the independent evaluation protocol and analysis to Section 5.5, with the main results summarized in Fig. 4. The complete model comparison and ablation results are provided in Tables 5 and 6 in Appendix B.2, respectively.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 3

Multi-reward optimization and weight sensitivity (Weakness 2 and Question 3)

We thank the reviewer for highlighting the need to compare alternative reward aggregation strategies and examine sensitivity to reward weights. We agree that both comparisons are important for assessing the effectiveness and robustness of multi-reward optimization. We address these two aspects below.

  1. Comparison of reward aggregation strategies. The w/o DGN ablation in Table 4 (Section 5.4) implements weighted-sum normalization (WSN). It first aggregates the raw IA, BP, and QP rewards using weights of 0.45, 0.30, and 0.25, and then normalizes the resulting scalar reward within each prompt group. In contrast, the full model first normalizes each reward dimension separately within its prompt group and then combines the normalized values using the same weights. Both configurations retain the subsequent batch normalization and advantage clipping. This comparison therefore tests aggregation before versus after dimension-wise normalization, with the reward weights held fixed.

    As shown in Table 4, under Gemini-3.1-Flash-Lite, w/o DGN obtains higher IA than the full model (8.148 versus 7.950), but lower BP (7.433 versus 8.264), QP (7.432 versus 7.631), and Overall (7.755 versus 7.964). The full design thus achieves a better balance across the three editing objectives.

    Independent evaluation in Table 6 (Appendix B.2) further supports this comparison. The full model exceeds w/o DGN in all three dimensions under both GPT-5.6-Terra and Gemini-3.6-Flash, with Overall increasing from 6.954 to 7.237 and from 5.436 to 6.065, respectively. These consistent gains support the benefit of DGN beyond the evaluator used for training rewards.

  2. Sensitivity to reward weights. We have completed additional experiments comparing the default weights with equal weighting and an IA-emphasized configuration on VIGOR-his. All configurations are evaluated using Gemini-3.1-Flash-Lite, with fixed Overall weights of (0.45, 0.30, 0.25). The results are reported in the table below.

    ConfigurationTraining weights (IA, BP, QP)IA ↑BP ↑QP ↑Overall ↑
    Equal(1/3, 1/3, 1/3)7.8788.2417.7227.948
    Default(0.45, 0.30, 0.25)7.9508.2647.6317.964
    IA-emphasized(0.60, 0.20, 0.20)8.1108.0517.6117.967

    The table shows that Overall remains stable between 7.948 and 7.967, with both alternative configurations differing from the default by less than 0.21%. Increasing the IA weight improves instruction alignment, while equal weighting yields the highest QP. The default configuration achieves the highest BP and an Overall close to that of the IA-emphasized configuration, balancing accurate editing with preservation of surrounding structures. These results show that reward weighting adjusts the balance among editing objectives while maintaining stable overall performance, indicating that CrossTimeEdit is robust to the tested variations in training reward weights.

The reward aggregation comparison is reported in Table 4, with additional independent evaluation results in Table 6 in Appendix B.2. We have added the reward-weight sensitivity analysis to Section 5.6 and Fig. 5, and the complete numerical results to Table 7 in Appendix B.3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to Reviewers

We sincerely thank all reviewers for their careful assessment and constructive suggestions. We are encouraged by their recognition of the value of VIGOR-his in addressing the gap in historical cross-view observations (Reviewers 6znu, vm1q, d2xQ), the well-motivated editing formulation that incorporates temporal satellite evidence while preserving viewpoint and unchanged scene content (Reviewers 6znu, vm1q, d2xQ), and the comprehensive ablation studies examining the contributions of the framework’s components (Reviewers 6znu, vm1q, d2xQ).

Building on their feedback, we have strengthened the validation of dataset quality and historical editing performance, expanded the analysis of multi-reward optimization, clarified the dataset construction and experimental subsets, and further defined the scope of historical street-view generation. All revisions are marked in blue in the manuscript. The key updates are summarized below.

Summary of Updates

  1. Human validation of dataset quality (Reviewers 6znu, vm1q). We added a stratified human audit of 200 quadruplets covering all 11 cities and four change categories in Appendix B.1, with a summary in Section 3. Dataset categories agree with human judgments in 93.0% of cases. The audit also assesses spatiotemporal pairing, within-period state consistency, and instruction grounding, supporting the reliability of VIGOR-his and its editing annotations.

  2. Independent evaluation. (Reviewers 6znu, vm1q, d2xQ): Fig. 4 in Section 5.5 summarizes evaluations by GPT-5.6-Terra and Gemini-3.6-Flash, with complete results in Tables 5 and 6 in Appendix B.2. CrossTimeEdit ranks first among the evaluated open-source image editing models in IA, QP, and Overall under both evaluators, improving Overall over the pretrained FLUX.2 [Klein] 4B baseline by 3.50% and 7.76%, respectively. Independent ablations further confirm the gains from SFT to RL and the contributions of the framework’s components.

  3. Historical appearance ambiguity. (Reviewers 6znu, d2xQ): We expanded Appendix E to distinguish satellite-observable structural changes from historical facade, material, and occlusion details that satellite observations cannot uniquely determine. Appendix F discusses complementary archival photographs and temporally and geographically grounded textual records, together with uncertainty modeling and the generation of multiple plausible historical appearances.

  4. Task scope and qualitative interpretation (Reviewers 6znu, d2xQ): Section 4.1 clarifies that historical street-view generation focuses on persistent building and road changes, rather than the exact reconstruction of transient objects such as pedestrians and vehicles. Section 5.3 highlights that the qualitative results demonstrate CrossTimeEdit’s ability to produce visually plausible outputs while recovering the semantic layout of buildings and roads in the earlier street views.

  5. Screening statistics. (Reviewer vm1q): Appendix D now reports the numbers of candidates removed and retained at each screening stage, documenting how 76,039 candidate quadruplets yield the final 43,653 quadruplets in VIGOR-his.

  6. Dataset and experimental subsets. (Reviewer vm1q): Section 5.1 clarifies that the training, validation, and test splits contain 24,182 quadruplets, comprising all 10,986 changed samples and 13,196 of the 32,667 no-change samples in VIGOR-his.

  7. Reward-weight sensitivity. (Reviewer 6znu): We added Section 5.6 and Fig. 5 to compare the default IA/BP/QP weights (0.45, 0.30, 0.25) with equal weighting and an IA-emphasized configuration (0.60, 0.20, 0.20). Overall remains stable between 7.948 and 7.967 across configurations, while individual criteria reflect the corresponding optimization priorities. These results demonstrate robustness to the tested reward-weight choices. Complete results are provided in Table 7 in Appendix B.3.

  8. Cross-view comparison with the same backbone. (Reviewer d2xQ): We added a FLUX.2 [Klein] 4B baseline fine-tuned with LoRA on VIGOR-his for satellite-to-street generation. Using the same pretrained foundation model, CrossTimeEdit achieves higher CV-C, CV-VR, and CV-PQ, with Overall increasing from 1.892 to 3.546. The results are included in Table 3, with the baseline introduced in Section 5.1, the comparison updated in Section 5.2, and training and inference configurations documented in Appendix B.8.

Below, we provide point-by-point responses to each reviewer, with the relevant evidence and revision locations. We appreciate the reviewers’ suggestions, which have helped improve the evaluation and presentation of this work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.