FaithDiff
FaithDiff is a state-of-the-art image super-resolution and restoration framework presented at CVPR 2025, designed to unleash diffusion priors for faithful image enhancement. Developed by researchers at the IMAG Lab, Nanjing University of Science and Technology, it excels in rejuvenating classic film footage, reviving old photographs, restoring social media images, and enhancing AIGC content. The system addresses real-world degradation scenarios with unknown distortions, supporting ultra-high-resolution output up to 8K and beyond on consumer-grade 24GB GPUs. Key technical features include integration with Hugging Face Diffusers, support for FP8 inference, and CPU offloading to significantly reduce memory consumption. FaithDiff provides pre-trained models, training code, and a dataset called RealDeg containing 238 samples of degraded images for evaluation. It is suitable for applications requiring high-fidelity reconstruction without introducing artificial artifacts, making it ideal for digital archiving, media