GALOSH: Training-free denoiser beats deep learning models with 650x speedup
A new algorithm removes search entirely, runs on CPUs, and outperforms BM3D and NLM baselines.
GALOSH (Generalized Anscombe LOcal SHrinkage), introduced by Yoshiro Sato in a July 2026 arXiv paper, reimagines classical training-free denoising by eliminating search entirely. While traditional methods like BM3D rely on content-dependent block matching that parallelizes poorly and requires GPU memory, GALOSH uses a fixed computation graph for every pixel: a blind per-image Poisson-Gaussian noise fit, a generalized Anscombe transform, two-pass local Walsh-Hadamard shrinkage for luminance, and luminance-guided local regression for chrominance. This design is fully data-independent and regular, allowing deterministic latency and easy mapping to fixed-point and streaming hardware.
On four real-noise benchmarks (SIDD Medium, RawNIND, raw and sRGB), GALOSH consistently beats BM3D and NLM baselines — even when those baselines are given oracle noise levels — and approaches the quality of trained networks on raw data. Its search-free architecture yields massive speed gains: 7x–650x faster than deep learning methods on a GPU at full benchmark size, and it is the only strong method that runs practically on plain CPUs. An operation-count analysis and working INT16 fixed-point realization confirm its suitability for edge and embedded devices without GPU access.
- Training-free, blind denoising for raw Bayer mosaics and sRGB/YUV images with zero domain-specific tuning.
- 7x–650x faster than deep learning baselines on GPU, and functional on plain CPUs via INT16 fixed-point implementation.
- Outperforms BM3D and NLM family on SIDD and RawNIND benchmarks, approaching trained networks on raw data.
Why It Matters
Fast, hardware-friendly denoising that works without training data — ideal for embedded cameras, mobile devices, and real-time pipelines.