PMRF

Why we included this project

PMRF goes after a problem that shows up constantly in image restoration: you want reconstructions that look photo-realistic, but you also want them to stay faithful to the original image. The method, published at ICLR 2025, uses rectified flow to approximate the estimator that minimizes mean squared error under a perceptual quality constraint, and the official repo ships a trained blind face restoration checkpoint plus a Hugging Face demo, so you can test it on low-quality face images without training anything yourself. Beyond face restoration, the codebase covers classic inverse problems like denoising, inpainting, and colorization, with checkpoints for controlled comparisons against a plain MMSE baseline. The PyTorch Lightning structure keeps it approachable for researchers adapting the flow formulation to their own tasks, and the theoretical framing helps if you need to justify the perceptual quality versus fidelity trade-off. One caveat: the HDiT architecture depends on the natten package, which has been finicky to install on Windows.