JJLibra/SALAD-Pan

Leverages a two-stage pipeline combining Band-VAE for sensor-agnostic latent compression with adapter-based fine-tuning of Stable Diffusion, enabling cross-sensor pan-sharpening without retraining the base model. Built on PyTorch and Hugging Face Diffusers with distributed training via Accelerate, supporting efficient inference through xformers optimization. Achieves state-of-the-art results across multiple satellite sensors (WorldView-3, etc.) by operating in learned latent space rather than pixel space, generalizing beyond training datasets.