Introduction
HYPIR (Harnessing Diffusion‑Yielded Score Priors) is a research implementation for image restoration and super-resolution that uses diffusion model derived score priors to improve reconstruction quality. The repository provides pretrained weights, examples, and demo links for reproducibility.
Key Features
- Diffusion‑based score prior technique to boost image restoration and super‑resolution performance.
- Pretrained SD2-compatible weights, inference and training scripts, and demo integrations on Replicate/OpenXLab.
- Support for batch inference and configuration examples for reproducible experiments.
Use Cases
- Research reproduction and algorithmic development for image restoration tasks.
- Benchmarking and comparison of restoration approaches using provided baselines and weights.
- Quick demos and experimentation through Colab and hosted demo platforms.
Technical Highlights
- Implemented in Python with PyTorch; includes scripts for training and inference (batch support).
- Integrates with Stable Diffusion 2.x checkpoints and provides config templates for training and evaluation.
- Comprehensive documentation and examples; note the repository includes a non-commercial use declaration—please follow the stated terms.