Sharing a denoiser I’ve been working on that might interest raw shooters here. GALOSH is classical (no AI, no training data) and fully blind — it estimates the noise model from your file, so there are no profiles to pick and it doesn’t care what camera you use.
Quality-wise it consistently beats the BM3D/NLM family on the public benchmarks (SIDD, RawNIND) at much lower runtime, especially on perceptual metrics — there’s a side-by-side viewer with every benchmark image if you want to pixel-peep: GALOSH — synchronized image comparison
For trying it on your own files there are Windows standalone builds — drag a DNG onto the exe, get name_GALOSH_l1_c1.dng next to it with all metadata copied (denoised DNGs open normally in darktable/RawTherapee): Releases · luxgrain/GALOSH · GitHub
On Linux it builds from source in seconds — it’s plain C99 + OpenMP, no dependencies (cd standalone && make all; the OpenCL GPU path just needs the OpenCL headers/ICD). The drag-and-drop bundles are Windows-only so far — happy to help anyone who wants to script it against their raws.
Everything is Apache-2.0 and the algorithm is documented to reimplementation level. A native darktable integration is a direction I’m actively working on (the code was written darktable-idiomatic from the start — the core entry point literally takes a dt_iop_roi_t), targeting the raw-domain denoise stage before demosaic. Feedback from this community on where it helps and where it falls short would be really valuable — high-ISO chroma blotches were the original motivation.
looks really interesting, i might try your code in my vulkan command buffers too see how it compares against other approaches. quite a lot of compute shaders there! are these variants or do we require all of these for one evaluation?
