Fig. 10: Performance evaluation.
From: Conditional neural field latent diffusion model for generating spatiotemporal turbulence
a Inference time cost comparison among OpenFOAM ran on CPU and Diff-FLowFSI ran on GPU, Video-diffusion model40, and CoNFiLD (ours). Red and blue regions denote the computational cost for dynamics generation and decoding, respectively. b GPU memory cost comparison for unconditional generation with (red line) and without (gray line) CNF encoder. c GPU memory cost comparison for implicit (ours, denoted by red bars) and explicit decoding strategy (e.g., POD, CNN, denoted by gray bars) in subsampling-based conditional generation.
