CycleISP

Why we included this project

Training denoisers on synthetic Gaussian noise often leaves them struggling on real camera photos, because real noise is signal-dependent and shaped by the imaging pipeline. CycleISP tackles that by modeling the camera pipeline in both directions, converting clean sRGB images to RAW and back, so it can generate realistic clean/noisy pairs in either space. That lets you build training data for RAW and sRGB denoising from ordinary images instead of collecting expensive real noisy datasets. The repo includes pretrained models and evaluation scripts for the DND and SIDD benchmarks, so you can reproduce the paper's results or retrain on your own data. It's a research codebase, so expect to work directly with PyTorch and the provided scripts.