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

Hanwen Lu; Jun He; Mingjia Yang; Hao Wei; Jinhao Huang; Yi Lin; Xiang Zhang

cs.CVSep 29, 2026

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

Authors: Hanwen Lu, Jun He, Mingjia Yang, Hao Wei, Jinhao Huang, Yi Lin, Xiang Zhang

Organizations: Sun Yat-sen University

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://luhanwen67.github.io/CrossTimeEdit-release/.

Figures & tables

Figure 2: VIGOR-his dataset construction pipeline, including data preprocessing, viewpoint screening and change classification, cross-view and temporal consistency screening, satellite-based change description generation, local editing instruction generation, and instruction consistency validation. Temporal satellite differences guide instructions describing the earlier states of changed regions, while recent street views anchor viewpoint and unchanged appearance.

Dataset	Cross-view	Temporal	Temporal	Cross-view	Text	Continents	Size
Dataset	pairing	street views	satellite views	time alignment	modality	Continents	Size
CVUSA	✓	✗	✗	✗	✗	1	44,416 pairs
CVACT	✓	✗	✗	✗	✗	1	128,334 pairs
VIGOR	✓	✗	✗	✗	✗	1	105,214 panoramas
CVGlobal	✓	✓	✗	✗	✗	6	134,233 pairs
CV-Cities	✓	✗	✗	✗	✗	6	223,736 pairs

Table 1: Comparison with related cross-view and temporal datasets. ✓: provided; ✗: not provided in the cited release.

Table 3

Figure 3: Qualitative comparison on five VIGOR-his cases. For readability, the satellite-based instructions shown in the figure are simplified for visualization.

Method	IA ( $\uparrow$ )	BP ( $\uparrow$ )	QP ( $\uparrow$ )	Overall ( $\uparrow$ )
Baseline	5.960	8.008	6.861	6.800
w/o RL	6.919	6.698	6.543	6.759
w/o SFT	6.085	8.112	6.886	6.893
w/o GT	7.938	7.356	7.508	7.656
w/o DGN	8.148	7.433	7.432	7.755
Full	7.950	8.264	7.631	7.964

Table 4: Ablations of CrossTimeEdit under the same VLM protocol.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Phase 1	Phase 2
Backbone	FLUX.2 [Klein] 4B; bf16 weights
Trainable modules	Attention-only LoRA; rank 32, alpha 32; text-stream projections included
Frozen modules	Base transformer, MLPs, text encoder, and VAE
Data	8,797 change + 13,196 no-change; shuffled; one pass per epoch
Epochs	0–4 (5 epochs)	5–9 (5 epochs)
Initialization	Pretrained backbone	Own epoch-4 LoRA checkpoint

Appendix

Table 5: Complete supervised fine-tuning configuration.

Setting	Value
Initialization / reference	SFT epoch 9 / frozen SFT epoch 9
Trainer / aggregation	Flow-GRPO; DGN, weighted aggregation, then batch normalization
Group sampling	50 conditions per round; 8 candidates per condition; 2 updates per round
LoRA / precision	Rank 32, alpha 32; fp32 master weights; bf16 mixed precision
Optimizer	AdamW; $7.5\times 10^{-6}$ ; weight decay $10^{-4}$ ; $\beta=(0.9,0.999)$ ; $\epsilon=10^{-8}$
Gradient norm / RL seed	1.0 / 42

Appendix

Table 6: Complete Flow-GRPO configuration for CrossTimeEdit.

Figure 4: Main-experiment training rewards and validation scores. (a) Mean IA, BP, and QP rewards over valid candidates at each RL epoch from 1 to 240. (b) Mean IA, BP, and QP scores on the full 1,095-location change validation split for the pretrained Baseline, SFT checkpoints e4 and e9, and RL checkpoints e20–e240.

Model	Human win rate (%)	Human rank (11)	Editing Overall	Editing rank (8)	Cross-view Overall	Cross-view rank (4)
Qwen-Image-3.0-Pro	80%	1	8.042	3	–	–
GPT-Image-2	72%	2	8.581	1	–	–
Nano Banana 2 Lite	72%	2	8.302	2	–	–
CrossTimeEdit	64%	4	7.964	4	3.546	1
FLUX.2 [Klein] 4B	58%	5	6.800	7	–	–
LongCat-Image-Edit	58%	5	7.360	6	–	–

Appendix

Table 7: Human preference and VLM evaluation results. Human win rates and human ranks use 50 comparisons per model across all 11 models. Editing rank and Cross-view rank denote the rankings by Editing Overall among the eight general image editing models and by Cross-view Overall among the four cross-view generation models, respectively. A dash denotes an inapplicable protocol.

Figure 5: Comparison between human preference and VLM evaluation. Left: human win rate versus Editing Overall for the eight general image editing models. Right: human ranks compared with Editing Overall ranks for the eight general image editing models (upper group) and Cross-view Overall ranks for the four cross-view generation models (lower group).

Figure 6: Representative VIGOR-his samples and their complete local editing instructions. Each sample contains a temporal cross-view quadruplet of co-located earlier/recent satellite images and street-view panoramas. From top to bottom, the examples cover no change (two locations), building changes, road changes, and combined building and road changes. The instructions organize the panorama by viewing direction and specify region-level ‘MODIFY’ or ‘PRESERVE’ operations.

Figure 7: Qualitative ablation comparison on three matched locations. Each row shows earlier and recent satellite images, the recent street view, the earlier street-view target, and outputs from the Baseline, w/o SFT, w/o GT, w/o DGN, w/o RL, and Full variants. For readability, the satellite-based instructions shown in the figure are simplified for visualization.

Figure 8: Qualitative comparison with cross-view generation models. Each row shows an earlier satellite image, outputs from Sat2Density and ControlS2S, the CrossTimeEdit output, and the corresponding earlier street-view ground truth.

Figure 9: Failure cases of historical street-view generation. The top case illustrates limited street-level visibility and a temporal state mismatch within the earlier cross-view pair; the bottom case illustrates ambiguity when generating a complex historical building from limited satellite evidence. For readability, the satellite-based instructions shown in the figure are simplified for visualization.

Explore similar work

StreetDiff: Multi-view Street Scenes Generation via Cross-view Consistent Multi-view Stable Diffusion with Structure Prompts
Sep 9, 2026Qi Zhang, Yanyifan Wang, Weiyuan Zhang +1Cross-View Video Diffusion Models
College of Computer Science and Software Engineering, Shenzhen University, China
Designing streetscapes from street-view imagery using diffusion models
May 17, 2026Yuzhou Chen, Yuebing Liang, Lingqian Hu +4Urban Environments Indoor Scene Generation
Department of Urban and Regional Planning, University of Florida · Singapore-MIT Alliance for Research and Technology Centre (SMART) · Department of Landscape Architecture and Urban Planning, Texas A&M University +1
You Cannot Photograph the Same Street Twice: Reliability Limits in Vision-Language Measurement of Urban Change
Sep 1, 2026Kaizhen TanUrban Environments Vision-Language Foundation Models
New York University

Sep 9, 2026, cs.CVJ/K move · Enter open · S save

StreetDiff: Multi-view Street Scenes Generation via Cross-view Consistent Multi-view Stable Diffusion with Structure Prompts

Qi Zhang, Yanyifan Wang, Weiyuan Zhang, Hui Huang

College of Computer Science and Software Engineering, Shenzhen University, China

Cross-View Video Diffusion Models Multimodal Diffusion Models Spatial Attention Prompt Engineering

Open paper AskPDF arXiv

Multi-view diffusion models have shown strong performance in scenes with strong geometric priors and sparse semantics, such as indoor rooms or simple outdoor environments (e.g., fields, courtyards). However, they often fail to maintain cross-view consistency under camera rotation, especially in structurally complex urban environments. Without explicit modeling of spherical correspondence across views, existing approaches tend to produce object duplication, structural distortion, and layout inconsistency. To address this limitation, we propose StreetDiff, a multi-view diffusion framework that explicitly enforces cross-view alignment during denoising. StreetDiff introduces a Panorama--Perspective Synergy design to decouple global layout reasoning from local detail synthesis, and incorporates a Panorama Alignment Module (PAM) that establishes spherical-projection-based attention constraints across views. By injecting structured alignment constraints without modifying the diffusion backbone, our framework achieves robust cross-view coherence in challenging urban street scene generation tasks. In addition, we construct Street360, a large-scale HDR multi-view urban panorama dataset. Extensive experiments demonstrate that StreetDiff significantly improves structural consistency and visual fidelity compared to prior multi-view diffusion generation methods.

Abstract

Explore similar work

StreetDiff: Multi-view Street Scenes Generation via Cross-view Consistent Multi-view Stable Diffusion with Structure Prompts

Designing streetscapes from street-view imagery using diffusion models

You Cannot Photograph the Same Street Twice: Reliability Limits in Vision-Language Measurement of Urban Change