ohayonguy/PMRF
A diffusion-model-based image restoration algorithm that approximates the optimal MSE-optimal estimator under perceptual quality constraints.
Velocity · 7d
+1.2
★ / day
Trend
→steady
star history
Official implementation of Posterior-Mean Rectified Flow (PMRF), an ICLR 2025 paper on photo-realistic image restoration. The method uses rectified flow matching with diffusion models to restore degraded images (denoising, inpainting, colorization, face restoration) while minimizing mean squared error under perceptual quality constraints. Built with PyTorch and Lightning, utilizing HDiT model architecture.