CrossTimeEdit: A Decade-Spanning Cross-View Dataset and Reward-Guided Editing for Historical Street-View Generation
CrossTimeEdit treats historical street-view generation as a constrained edit, using current street views and satellite change evidence to fill gaps in the past.
The paper introduces VIGOR-his, a dataset of 43,653 location-level quadruplets across 11 cities on three continents. Its pipeline pairs views, screens consistency, classifies changes, and builds satellite-based editing instructions. The model starts from FLUX.2 [Klein] 4B, then uses supervised fine-tuning and online reinforcement learning. The authors report a 17.12% overall gain over the pretrained baseline across instruction alignment, background preservation, and visual plausibility. HF Daily Papers' note
The paper introduces VIGOR-his, a dataset of 43,653 location-level quadruplets across 11 cities on three continents. Its pipeline pairs views, screens consistency, classifies changes, and builds satellite-based editing instructions. The model starts from FLUX.2 [Klein] 4B, then uses supervised fine-tuning and online reinforcement learning. The authors report a 17.12% overall gain over the pretrained baseline across instruction alignment, background preservation, and visual plausibility. HF Daily Papers' note
score 4