AI raw denoise
Removes sensor noise from the raw mosaic with a neural network, before white balance, chromatic-aberration correction and demosaicing — so every later stage of the pipeline inherits clean data, and demosaicing interpolates real detail instead of weaving noise into maze and zipper artifacts.
Unlike general-purpose AI denoisers, these models are trained specifically on non-demosaiced RAW sensor data, for the exact point of the Ansel pipeline where the module runs, using Ansel’s per-camera noise profiles to synthesize physically accurate Poisson-Gaussian noise. One set of weights covers every camera Ansel has a noise profile for, Bayer and X-Trans alike — a newly profiled camera is supported without retraining. The models weigh between 7 and 36 MB each and run entirely locally, on CPU (OpenMP) or GPU (OpenCL); nothing ever leaves your machine.




