The authors added a second prediction stream to a pretrained video diffusion model.
Predicting velocity fields helps video models generate more realistic fluids
FluidGen post-trains a video diffusion model to generate RGB frames and three-dimensional velocity fields jointly, then uses the predicted motion to favor physically plausible outputs.
Top University
Yang Zheng · Zizhang Li · Boyang Deng · Jan Ackermann · Lei Shu · Yinghao Xu · +3 more
Stanford University
Research Digest··2 min read
Zheng, Li, Deng and colleagues target a persistent weakness of video generators: smoke, fire and explosion debris often look detailed but move without coherent turbulence.
§