Reproduce “sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongjin Kim, Donggoo Jung, Sungyong Baik, Tae Hyun Kim
NAFlow, a Normalizing Flow-based framework, effectively models diverse real sRGB image noise distributions, including spatially correlated noise, through a unified model and a Noise-Aware Sampling algorithm that eliminates metadata dependency. Training a standard denoising network with NAFlow's synthetic noise achieved 37.22 dB PSNR on SIDD-Benchmark, outperforming previous synthetic noise generation methods and significantly reducing the gap to real-noise training.
Key insights
- Real-world sRGB noise from diverse camera settings can be effectively modeled by associating each configuration with a distinct Gaussian distribution in the latent space of a Normalizing Flow.
- Noise-Aware Sampling allows a single unified model to implicitly infer and generate appropriate noise characteristics for new noisy images, even without explicit camera metadata.
- Capturing multi-scale spatial correlations, often induced by camera ISPs, is essential for generating highly realistic sRGB noise, which can be achieved through multi-level architectural designs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.