GeoCR: one generalist model unifies cloud removal across 10 datasets, no fine-tuning needed
KAIST-VIClab · hf · 2026-10-06
KAIST-VIClab releases GeoCR, a generalist cloud removal prior for remote sensing.
- Problem: existing CR methods are specialized to individual datasets and input configurations, limiting reuse across sensors, spectral bands, and observation settings.
- Method: compact input/output stems extend a pretrained RGB autoencoder (trunk frozen), letting a single flow transformer jointly model clean RGB and non-RGB latents from single/multi-temporal cloudy observations, with optional SAR tokens.
- Training: joint pretraining on 10 datasets with 883,331 cloud-free targets; the same checkpoint supports direct inference without dataset-specific fine-tuning and LoRA adaptation.
- Results: best FID and DISTS on full-band SEN12MS-CR and Sen2MTCNew and RGB-only CUHK-CR2, outperforming existing CR and restoration models.
More from Research
- LLM-assisted 22-page preprint settles Hilbert's 12th problem and Zauner's conjecture — basedjensen · 2026-10-06
- 18,000+ runs: new study shows how agent configuration shapes performance on 4 scientific tasks — zeynepakata · 2026-10-06
- Marin team heads to COLM to discuss data, scaling laws, and open source — dlwh · 2026-10-06
- Hamel Husain: similarity metrics like ROUGE don't work for LLM output evals — HamelHusain · 2026-10-06
- Nearly half of DOL-recognized jobs have zero agentic AI tool coverage, Cohere finds — Cohere_Labs · 2026-10-06
- Princeton-affiliated paper: stochastic contexts make adversarial online learning efficiently computable — HazanPrinceton · 2026-10-06