The project introduces InvSR, an image super-resolution technique that leverages pre-trained diffusion models as priors. It proposes a Partial Noise Prediction strategy to construct an intermediate state of the diffusion model, which serves as a flexible starting point for sampling. The core component is a deep noise predictor that estimates optimal noise maps for the forward diffusion process, enabling high-quality super-resolution results in 1 to 5 sampling steps.