<ul data-eligibleForWebStory="true"><li>This work introduces nonnegative diffusion (NnD) as a generative model for creating computationally tractable, physically meaningful, and costly-to-simulate objects.</li><li>NnD uses score-based diffusion and annealed Langevin dynamics to ensure non-negativity during scene generation and analysis.</li><li>The model is trained on high-quality physically simulated objects and can be used for generation and inference after training.</li><li>The researchers demonstrate the generation of 3D volumetric clouds using NnD, which exhibit characteristics consistent with cloud physics trends and are difficult for experts to distinguish as non-physical.</li></ul>

NnD: Diffusion-based Generation of Physically-Nonnegative Objects

Discover more