Fluid-Gen-Zero hands water physics to a simulator and cuts fluid flow error by 84%

Researchers have published Fluid-Gen-Zero on arXiv, a video generation framework in which objects interact with fluid. It requires no fine-tuning. On three backbones (Tora, built on CogVideoX, plus VACE and WanMove, both built on Wan), object trajectory error falls by 26.7–81.5%, and fluid flow endpoint error (fEPE) falls by 67.9–84.0%. Visual quality is largely preserved.

Anyone who has asked a video model to show a ball dropping into water knows the problem. The footage looks convincing, but the splashes, the waves and the ball's trajectory each follow their own rules. The model learns from pixels and only imitates physics. The authors split the work: a physics simulator computes the motion, while a pretrained generator still renders the appearance.

Two levels of the agentic pipeline

An agentic workflow with two stages connects the two worlds. The first is called generation-time planning. An agent built on a vision-language model reads the user's intent and the simulation output, then breaks the generation into clips. The second stage, latent-space guidance, works at a deeper level. Simulation signals are injected into the denoising process through region-aware latent wrapping, meaning the latent representations are adjusted to match the simulated motion in the relevant regions of the frame.

The video model's weights stay untouched. The authors describe the scheme as plug-and-play and claim it is compatible with modern video foundation models. The key bet is that physics doesn't need to be baked into a neural network through expensive training if a simulator can supply it and the model can take it as a hint.

To test the approach, the team built its own benchmark for video generation involving fluid-object interaction.

What the measurements show

• Object trajectory error: a reduction of 26.7–81.5%, depending on the backbone.
• Fluid flow error (fEPE): a reduction of 67.9–84.0%.
• Comparison with the same model without the method: in a human survey, raters chose Fluid-Gen-Zero in 55.1–74.4% of cases across the three backbones.
• Comparison with simulation-based methods: preferred in 90.4–94.2% of cases.
• Perceptual quality: preserved "to a large extent," which means small losses are possible.

The spread in the numbers is wide. The low end of the trajectory improvement (26.7%) is noticeably more modest than the high end, and the 55.1% preference on one of the backbones is only slightly above random choice. The method wins by the largest margin against purely simulation-based competitors: they produce correct physics, but their appearance falls short of video generators.

The caveats are obvious. This is a preprint without peer review, the benchmark is the authors' own, and the code and data are promised only after the paper is accepted. The abstract says nothing about compute cost or generation time. It also leaves open how much the method depends on the accuracy of the scene defined in the simulator.

Where it could be useful

The most likely early users are visual effects studios and advertising teams that need believable water without assembling the scene by hand in a simulator and without rendering from scratch. A second niche is synthetic data for robotics and world models, where correct fluid behavior is critical. For product teams, something else matters more: the scheme works on top of models they already use, so there is no need to retrain the stack.

It is worth waiting for the code and checking three things: behavior on complex scenes with several objects, the cost of wiring the simulator to the VLM agent, and robustness on models that are not on the authors' list. If the results reproduce, physical grounding through an external simulator could become a cheap alternative to teaching physics inside the video model.