Sketch-guided flow field generation with diffusion model
Hengyuan Chang, Yichen Peng, Syuhei Sato, Haoran Xie · 2024
The convergence of flow field generation with deep learning for augmenting the efficiency of flow simulation is a prominent pursuit. Current approaches primarily utilize conditional generative adversarial networks(cGANs) to generate velocity fields guided by sketches. In contrast, the cGAN training process exhibits instability. In this study, we propose a novel 2D velocity field design and generation framework that leverages the latent diffusion model (LDM). The sketch is a constraining condition that guides the denoising process within LDM and 2D velocity field reconstruction. Our framework is proficient in generating velocity fields that align with the shape of given sketches. We verified the robustness of the proposed framework in comparison to cGAN-based methods.