Kernekoncepter
A single generative model that can learn to synthesize a wide range of procedural noise patterns, including spatially-varying blends between them.
Resumé
The paper presents a method for learning a unified generative model that can produce a diverse range of procedural noise patterns, including spatially-varying blends between them. The key contributions are:
A denoising diffusion probabilistic model (DDPM) with a spatially-varying conditioning mechanism, which allows the model to generate noise patterns with spatially-varying characteristics.
A novel data augmentation strategy using CutMix, which enables the model to learn to respond to spatially localized conditioning signals, despite only having access to globally uniform noise samples during training.
Demonstration of the model's ability to synthesize a variety of visually compelling noise textures, including seamless tileable patterns and spatially-varying blends between different noise types.
An application of the model to inverse procedural material design, where it is used as a differentiable proxy for noise generator nodes in a material graph, improving the fidelity of material reconstructions.
The model is trained on a dataset of 1.2 million noise images covering 18 unique noise functions, with dense sampling of their respective parameter spaces. The authors show that their approach outperforms prior work in terms of both qualitative and quantitative metrics.
Statistik
The paper does not provide any specific numerical data or statistics to support the key claims. The results are primarily presented through qualitative visualizations of the generated noise patterns.
Citater
"Our method enables the synthesis of a wide range of noise patterns with spatially-varying characteristics."
"Our model creates semantically meaningful interpolations between noise configurations; above we see the Siggraph logo written with hay fibers that are nested inside of Damascus steel striations – the scale and distortion of the steel pattern naturally interpolates into a denser pattern before transitioning into fibers."