Hugging Face Releases CrossTimeEdit: Revolutionizing Historical Street-View Generation

Summary:Hugging Face has released CrossTimeEdit, a model that leverages recent street views and temporal satellite data to generate plausible historical street scenes. Built on FLUX.2 4B, the model employs a combination of supervised fine-tuning (SFT) and online reinforcement learning (RL), with three editing criteria: Instruction Alignment, Background Preservation, and Quality and Physical Plausibility. Experiments demonstrate that CrossTimeEdit outperforms existing cross-view generation models in scen

---

Technical Mechanism Analysis

The core of CrossTimeEdit lies in redefining the historical street-view generation task as an editing task. Its main technical features include:

1. Dataset Construction: CrossTimeEdit is built on the VIGOR-his dataset, which contains 43,653 location-level quadruplets across 11 cities on three continents. The dataset includes an automated pipeline for spatial pairing, consistency screening, change classification, and the generation and validation of satellite-based change descriptions and local editing instructions.

2. Model Architecture: CrossTimeEdit is trained based on the FLUX.2 4B model, using a combination of supervised fine-tuning (SFT) and online reinforcement learning (RL).

3. Editing Criteria and Reward Mechanism: The model defines three editing criteria: Instruction Alignment (IA), Background Preservation (BP), and Quality and Physical Plausibility (QP). These criteria also serve as reward signals for reinforcement learning and are optimized using Group Reward-Decoupled Normalization Policy Optimization (GDPO).

4. Performance Improvement: CrossTimeEdit improves overall performance across the three editing criteria by 17.12% over the pretrained baseline and outperforms existing cross-view generation models in scene consistency, visual realism, and perceptual quality.

Engineering Trade-offs and Actual Performance

• Advantages: CrossTimeEdit effectively combines recent street views with temporal satellite data to generate images with high visual realism and consistency. Its multi-reward optimization mechanism ensures excellent performance across multiple evaluation dimensions.

• Challenges: Although CrossTimeEdit performs well in many aspects, its training process requires large amounts of high-quality annotated data, and the reinforcement learning stage incurs significant computational overhead. Additionally, the model may still face challenges in handling extreme change scenarios, such as large-scale building demolition or construction.

Developer Implementation and Deployment Recommendations

• Data Requirements: Developers need to prepare high-quality recent street views and historical satellite imagery data, ensuring data consistency and completeness.

• Computational Resources: Due to the reinforcement learning training method, it is recommended to deploy the model on platforms with strong computational resources, such as GPU-accelerated servers.

• Application Scenarios: CrossTimeEdit is suitable for applications such as historical street-view reconstruction, urban planning, and cultural heritage preservation. Developers can further fine-tune or extend the model based on specific needs.

Conclusion

CrossTimeEdit represents Hugging Face's latest breakthrough in the field of historical street-view generation. Its innovative editing task definition and multi-reward optimization mechanism provide a new technical path for the field. Despite some challenges in data requirements and computational resources, its excellent performance makes it widely applicable in multiple scenarios.

---

Source: Hugging Face Daily Papers (2026-09-29)

— END —

Tags: #Hugging Face Follow #Historical Street-View Generation Follow #Reinforcement Learning Follow #Multimodal Follow #Cross-View Follow