ohayonguy/PMRF

Implements a rectified flow-based diffusion model with HDiT architecture that provably approximates the optimal MSE-minimizing estimator under perceptual quality constraints, using PyTorch Lightning for training. The approach combines diffusion-based generative modeling with theoretical guarantees on reconstruction quality, enabling configurable inference steps (K parameter) to trade computational cost against restoration fidelity. Integrates with standard vision benchmarks (CelebA, WIDER, WebPhoto) and provides pre-trained checkpoints via Hugging Face for blind face restoration tasks.

742 stars. No commits in the last 6 months.