CrossTimeEdit: Decade-Spanning Dataset and Reward-Guided Historical Street-View Generation
SunYatsen · hf · 2026-10-13
Researchers release VIGOR-his, a decade-spanning cross-view dataset with 43,653 location-level quadruplets across 11 cities on three continents, plus CrossTimeEdit, a model that reframes historical street-view generation as editing.
- Dataset: automated pipeline for spatial pairing, consistency screening (Gemini-assisted), change classification, and satellite-based editing instructions.
- Model: fine-tuned from FLUX.2 [Klein] 4B with SFT then online RL, optimizing three rewards (Instruction Alignment, Background Preservation, Quality/Physical Plausibility) via Flow-GRPO with GDPO per-reward normalization.
- Results: +17.12% overall over the pretrained baseline, outperforming GPT-Image-2, LongCat-Image-Edit and cross-view generation baselines on consistency, realism and perceptual quality. Code, data and weights are open-sourced.
More from Multimodal
- Let Gemma write your prompts: MiniMax H3 image-to-video pipeline with 2K upscale — EzkaProductions · 2026-10-13
- MiniMax H3 + Seedance Demo: Fast 3D White-Model Generation from Video Characters — Only_Airport7253 · 2026-10-13
- Recraft demos a full AI avatar pipeline: V4.1 Pro, Nano Banana, Seedance 2.5 — aziz4ai · 2026-10-13
- NeurIPS paper: video diffusion models encode physics internally even when outputs are wrong — katjahofmann · 2026-10-13
- Open-source .char format keeps characters consistent across video, voice included — ashishsanu · 2026-10-13
- AI-generated video of an abandoned 2036 London street goes viral on Reddit — ciphernull7166 · 2026-10-13