GeoCR: Learning a Generalist Cloud Removal Prior from Heterogeneous Observations
GeoCR is pitched as one checkpoint for cloud removal across RGB, multispectral, single- or multi-temporal inputs, with optional SAR guidance.
The model keeps a pretrained RGB autoencoder’s main encoder and decoder frozen, adding compact stems for different sensing domains. Its shared latent setup lets one flow transformer model clean RGB and non-RGB outputs from cloudy observations. The authors pretrain it on ten datasets with 883,331 cloud-free target images, then report direct inference and LoRA adaptation without dataset-specific full fine-tuning. They say GeoCR posts the best FID and DISTS on selected full-band and RGB-only benchmarks. HF Daily Papers' note
The model keeps a pretrained RGB autoencoder’s main encoder and decoder frozen, adding compact stems for different sensing domains. Its shared latent setup lets one flow transformer model clean RGB and non-RGB outputs from cloudy observations. The authors pretrain it on ten datasets with 883,331 cloud-free target images, then report direct inference and LoRA adaptation without dataset-specific full fine-tuning. They say GeoCR posts the best FID and DISTS on selected full-band and RGB-only benchmarks. HF Daily Papers' note
score 4