Reproduce “One Noise to Rule Them All: Learning a Unified Model of Spatially-Varying Noise Patterns” 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
Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir G. Kim, Matthew Fisher, Sören Pirk, Daniel Ritchie
A unified generative model synthesizes diverse procedural noise types, enabling smooth interpolation and spatially-varying patterns that overcome limitations of discrete noise selection and poor blending. This Denoising Diffusion Probabilistic Model, developed by Brown University and Adobe Research, achieves a mean FID of 20.9 across 18 noise types, improving upon previous methods.
Key insights
- Spatially-Adaptive Denormalization (SPADE) enables spatially-varying noise characteristics at inference time, even when trained on spatially uniform conditioning data.
- A modified CutMix data augmentation strategy trains the model to respond precisely to localized conditioning signals, facilitating spatial control over noise properties.
- Constraining class embeddings to a sphere improves interpolation quality between different noise types, allowing for smoother blending via spherical linear interpolation (slerp).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.