EndlessSora/focal-frequency-loss

Computes an adaptive frequency-domain loss that down-weights easy-to-synthesize frequency components while focusing on harder-to-reconstruct ones, complementing spatial losses in generative models. Implemented as a drop-in PyTorch `nn.Module` with tunable hyperparameters (`alpha` for focus strength, `loss_weight` for contribution), it integrates seamlessly with VAEs, pix2pix, SPADE, and StyleGAN2 architectures. Operates via Fourier transform analysis with configurable patch-based and batch-level spectrum weighting strategies across diverse image tasks from reconstruction to conditional and unconditional synthesis.

706 stars and 2,572 monthly downloads. No commits in the last 6 months. Available on PyPI.