FluidGen: Physics-Aware Fluid Video Generation via Motion-Guided Post-Training

Yang Zheng; Zizhang Li; Boyang Deng; Jan Ackermann; Lei Shu; Yinghao Xu; Jiajun Wu; Gordon Wetzstein

cs.CVOct 19, 2026

FluidGen: Physics-Aware Fluid Video Generation via Motion-Guided Post-Training

Authors: Yang Zheng, Zizhang Li, Boyang Deng, Jan Ackermann, Lei Shu, Yinghao Xu, Jiajun Wu, Gordon Wetzstein,

Organizations: Stanford University

Abstract

Video generation models produce visually compelling motion but systematically fail on fluid phenomena, including smoke, fire, and explosion debris, where outputs lack coherent vortical structure and collapse to bland dynamics that violate the physics of real fluids. Classical solvers achieve high physical fidelity but require detailed 3D scenes and carefully tuned parameters, making them impractical for in-the-wild imagery and video generation. We address this gap by injecting a data-driven fluid prior into a pre-trained video model, trained on a mixture of high-fidelity Houdini simulation data with paired velocity-field ground truth and curated in-the-wild fluid video data for appearance diversity. The core of our approach is a dual stream architecture that predicts per-frame velocity fields alongside RGB, supervised by simulation ground truth. This auxiliary signal regularizes generation toward physically consistent motion and exposes an interpretable physical quantity that we further leverage as a reward for velocity-aware direct preference optimization. Our model produces substantially more realistic fluid dynamics sequences than strong video baselines and optimization-based methods. Project page: https://y-zheng18.github.io/FluidGen/

Figures & tables

Figure 1: We introduce FluidGen , a physics-aware video diffusion model for realistic fluid dynamics. FluidGen jointly predicts per-frame velocity fields (insets) that regularize generation toward physically consistent motion.

Figure 2: Overview of FluidGen . Our pipeline augments a pretrained video diffusion model with a velocity prediction stream and train it in two stages: joint supervised fine-tuning on simulated and in-the-wild fluid videos, followed by velocity-aware DPO using physics-informed rewards.

Figure 3: Data visualization. Here we show two examples of simulated scenes: the rendered RGB frames, velocity field frames and visualized velocity directions in our Houdini-constructed dataset.

Figure 4: Physics-informed reward distinguishes plausible from implausible fluid dynamics. Four samples (left: RGB frame; right: visualized velocity field) from the same input annotated with their reward values. High reward samples exhibit organized, turbulent flow structure with rich directional diversity and coherent vortical motion, while low reward samples show collapsed velocity fields that do not correspond to realistic fluid behavior.

Figure 5: Qualitative comparison on in-the-wild inputs . Wan2.1-I2V-14B produces visibly unnatural fluid motion with appearance drift, while the model after the Joint Supervised Fine-tuning (FluidGen w/o DPO) generates physically coherent dynamics. Our full solution with DPO (FluidGen) further sharpens the result, yielding richer fine-scale details and more vivid fire and smoke motion. Please refer to our project page for better visualization.

Figure 6: Qualitative comparison on the FluidNexus ( Gao et al., 2025 ) dataset . Even with costly multi-view reconstruction, FluidNexus still omits some of the fine details, while the pretrained Wan2.1 fails to generate realistic fluid dynamics and consistent scene appearance.

Figure 7: Generated velocity fields on a real-world capture. Here we show the input RGB ( left ), the generated velocity encoded as color ( middle , Eq. 1 ), and the decoded velocity overlaid as arrows ( right ). The recovered field reveals coherent advection consistent with the observed smoke motion.

Figure 8: Fluid simulation from generated velocity fields . Our generated fluid velocity fields are physically accurate and can be directly employed in Houdini to simulate a fluid source.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Visual effects generation using FluidGen. The two rows are ordered frames from a video generated by FluidGen. Given keyframes, FluidGen generates a realistic fluid dynamic transition, here ink in water forming the letter “A”.

Figure 10: FluidGen qualitative results . Each row contains frames from a video generated by FluidGen. FluidGen can generate realistic fluid, such as smoke, ink, and fire, in various scenarios.

Category	Name	Computing formula
Structural	$R_{\mathrm{div}}$	$\displaystyle\mathbb{E}_{t}\!\left[\tanh\!\Big(\tfrac{1}{3}\!\!\sum_{k\in\{x,y,z\}}\!\!\mathrm{Var}\,\hat{u}_{k}\big\|_{\mathcal{A}_{t}}\big/\kappa_{\mathrm{div}}\Big)\right]$
	$R_{\mathrm{tc}}$	$\displaystyle\mathbb{E}_{t}\!\left[\big[\mathrm{Pearson}\!\left(\nabla\hat{\mathbf{u}}_{t},\,\nabla\hat{\mathbf{u}}_{t+1}\right)\big]_{+}\right]$
	$R_{\mathrm{msd}}$	$\displaystyle\mathbb{E}_{t}\!\left[0.7\tanh(g_{1}/\kappa_{g})+0.3\,\exp\!\big(-\tfrac{1}{2}\log^{2}(g_{1}/g_{\|\mathcal{S}\|})\big)\right]$
Statistical	$R_{\mathrm{act}}$	$\displaystyle\exp\!\Big(-\tfrac{1}{2}\big(\bar{f}_{a}-\mu_{a}\big)^{2}/\sigma_{a}^{2}\Big)$
	$R_{\mathrm{tdb}}$	$\displaystyle\exp\!\Big(-\tfrac{1}{2}\big(\mathbb{E}_{t}\|R_{t+1}-R_{t}\|-\mu_{R}\big)^{2}/\sigma_{R}^{2}\Big)$
	$R_{\mathrm{lc}}$	$\displaystyle\sum_{k}\lambda_{k}\,s_{k}(e_{t}),\quad e_{t}=\mathbb{E}_{x,y}[\rho_{t}]$

Appendix

Table 3: The thirteen component rewards combined into the total

R(\tilde{\mathbf{z}})=\sum_{i}w_{i}R_{i}

used by the velocity-aware DPO objective. See Sec. B.1 for notation and definitions of

\mathcal{A}_{t}

\hat{\mathbf{u}}

e_{t}

S_{2}(r)

, and the scalar hyper-parameters.

	Fire	Smoke	Explosion
Solver	Houdini Pyro	Houdini Pyro	Axiom (GPU)
Buoyancy	$\mathcal{U}(0.15,0.3)$	$\mathcal{U}(0.5,1.0)$	$\mathcal{U}(0.1,1.0)$
Turbulence	$\mathcal{U}(0.8,2.0)$	$10^{\mathcal{U}(-0.5,-0.05)}$	$\mathcal{U}(0,10)$
Disturbance	$\mathcal{U}(3,15)$	$\mathcal{U}(0.1,0.2)$	$\mathcal{U}(0,1)$
Dissipation	$0.01$	$\mathcal{U}(0.005,0.05)$	$\mathcal{U}(0.005,0.05)$
Divergence scale	$\sim\!\mathcal{U}(0.10,0.15)$	$\sim\!\mathcal{U}(0.10,0.15)$	—

Appendix

Table 4: Per-asset randomization ranges for Houdini simulation data construction.

\mathcal{U}(a,b)

stands for uniform distribution and

\mathcal{N}(\mu,\sigma)

stands for Gaussian distribution.

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training
Oct 7, 2026Hong Huang, Yuqiu Liu, Chenyu You +3Video Generation Physics-Guided Video Generation
Simon Fraser University · Stony Brook University · Lawrence Berkeley National Laboratory +1
PhysFlow: Physics-Aware Optical Flow for Motion Controllable Video Generation
Sep 8, 2026Cong Wang, Hanxin Zhu, Yonglin Tian +5Optical Flow Physical Consistency in Video Generation
State Key Laboratory of Multi-modal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Zhongguancun Academy, Beijing 100094, China +2
Physics-Grounded Fluid Video Generation with a Simulation Dataset and Dual-Stream Optical-Flow Supervision
Jul 28, 2026Ruijie Su, Yuanzhi Liang, Xiaohua Xie +1Physical Consistency in Video Generation Image-to-Video Generation
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Institute of Artificial Intelligence, China Telecom (TeleAI), China · Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, Guangzhou, China

Oct 7, 2026J/K move · Enter open · S save

Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training

Hong Huang, Yuqiu Liu, Chenyu You, Daniel Martin+2

Simon Fraser University · Stony Brook University · Lawrence Berkeley National Laboratory+1

Video Generation Physics-Guided Video Generation Physical Consistency in Video Generation

Open paper AskPDF arXiv

We present Fluid-Gen-Zero, a training-free framework for physics-aware fluid-object interaction video generation that decouples physical reasoning from appearance synthesis. Our key insight is to delegate motion dynamics to a physics simulator while preserving the appearance modeling capacity of pretrained video generators. We bridge these two domains through a two-level agentic workflow: generation-time planning, where a vision-language model (VLM) agent interprets intent and the simulation rollout to organize generation clips, and latent-space guidance, which injects simulation signals into denoising through region-aware latent wrapping. This plug-and-play design is compatible with current video foundation models. We further introduce a benchmark for fluid-object interaction video generation. Across Tora (CogVideoX-based), VACE and WanMove (Wan-based), Fluid-Gen-Zero consistently improves simulation alignment, reducing object trajectory error by 26.7%-81.5% and fluid fEPE (fluid flow endpoint error) by 67.9%-84.0%, while largely preserving perceptual quality. In a human preference study, raters favor Fluid-Gen-Zero in 55.1%-74.4% of same-backbone comparisons across three backbones, and in 90.4%-94.2% of comparisons against simulation-based methods. Code and data will be released upon acceptance.

Abstract

Video Generation

Explore similar work

Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training

PhysFlow: Physics-Aware Optical Flow for Motion Controllable Video Generation

Physics-Grounded Fluid Video Generation with a Simulation Dataset and Dual-Stream Optical-Flow Supervision