Fig. 1: Overview of the proposed Conditional Neural Field Latent Diffusion (CoNFiLD) model.
From: Conditional neural field latent diffusion model for generating spatiotemporal turbulence
a Architectures: a FiLM-based CNF for encoding dynamic flow sequences into latent space, where the underlying distribution of the latent vectors is implicitly captured by learning reverse diffusion (denoising) processes. b Zero-shot generation: synthesizing new spatiotemporal flow fields with arbitrary length, either unconditionally or based on specific conditions (e.g., sparse sensor data), without the need for retraining.
