[IGARSS ORAL 2026] Generates synthetic image–mask pairs by denoising joint latents conditioned on a real input image (VAE) and a class prompt (CLIP), enabling targeted oversampling of minority classes (e.g., barren) for LoveDA semantic segmentation.
generative-model class-imbalance semantic-segmentation data-augmentation earth-observation domain-adaptation diffusion-model synthetic-dataset-generation conditional-generation distribution-control latent-diffusion controlnet long-tail-learning ratio-conditioned-generation dataset-balancing loveda
-
Updated
May 8, 2026 - Python