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

Hong Huang1, Yuqiu Liu1, Chenyu You2, Daniel Martin3, Chuhang Zou4†, Wuyang Chen1†
1Simon Fraser University   2Stony Brook University   3Lawrence Berkeley National Lab   4Meta Reality Labs
† Correspondence
Paper arXiv Code
Fluid-Gen-Zero brings simulation-grounded physical control to pretrained video generators without updating their parameters.
Training-freeSimulation-groundedPlug-and-play

Interactive Physics Control

Apply a force and watch the fluid respond. Each scene is driven by one external force — press the arrow where it acts.

Abstract

We present Fluid-Gen-Zero, a training-free framework for physics-aware fluid 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 rule-grounded VLM agent parses intent, observes simulation evidence, and organizes generation into physically feasible clips, and latent-space guidance, which injects simulation-derived signals into denoising via region-aware latent wrapping. This plug-and-play design is compatible with modern video models and improves both physical alignment and visual quality without training.

Comparison between training-based and training-free physics-aware video generation.
Training-based vs. training-free physics-aware video generation. Existing approaches rely on architecture-coupled post-training, while Fluid-Gen-Zero decouples physical reasoning from appearance synthesis and introduces physics awareness entirely at inference time.
Simulation
structured dynamics
Agentic planning
event-aware clips
Latent guidance
appearance synthesis
Fluid-Gen-Zero pipeline from input image and prompt through simulation, agentic planning, latent control, and video generation.
Fluid-Gen-Zero pipeline. Starting from an input image and text prompt, the method reconstructs a simulatable 3D scene, rolls out fluid-object dynamics, and guides a frozen video foundation model through generation-level planning and latent-level control.

Qualitative Results

Ten benchmark scenes. Pick a backbone to compare it against its own Fluid-Gen-Zero result, alongside the simulation, two physics-based methods and the training-free MotionCraft.

Simulation
PhysGen3D
WonderPlay
MotionCraft
VACE
Ours (VACE)

Bird's-eye view showing a gold coin resting calmly on an abstract field of crystal clear milk, creating subtle interference patterns.

Bird's-eye view where a fresh red pear rotates rapidly on top of a turbulent, storm-tossed sea of milky oil, flinging droplets outwards due to centrifugal force.

Direct overhead view of a black ceramic cube resting in a shallow pool of purple liquid.

Flat-lay style shot capturing a waxed coconut smashing into a heavy, moving mass of milky slime.

Flat-lay style shot where a matte rubber thick ring rotates rapidly on top of a stagnant, still pond of translucent teal paint, flinging droplets outwards due to centrifugal force.

Strict overhead view of a bright yellow pomegranate slowly drifting across a perfectly uniform, monochromatic field of liquid mercury.

Strict top-down view of a copper pyramid resting in a shallow pool of opaque white paint.

Strict top-down view of a wet passion fruit resting in a shallow pool of purple oil.

Vertical aerial shot of a stationary waxed lemon.

Bird's-eye view of a shiny green coconut partially submerged in a high-contrast pool of milky juice.

Simulation
PhysGen3D
WonderPlay
MotionCraft
Tora
Ours (Tora)

Bird's-eye view showing a gold coin resting calmly on an abstract field of crystal clear milk, creating subtle interference patterns.

Bird's-eye view where a fresh red pear rotates rapidly on top of a turbulent, storm-tossed sea of milky oil, flinging droplets outwards due to centrifugal force.

Direct overhead view of a black ceramic cube resting in a shallow pool of purple liquid.

Flat-lay style shot capturing a waxed coconut smashing into a heavy, moving mass of milky slime.

Flat-lay style shot where a matte rubber thick ring rotates rapidly on top of a stagnant, still pond of translucent teal paint, flinging droplets outwards due to centrifugal force.

Strict overhead view of a bright yellow pomegranate slowly drifting across a perfectly uniform, monochromatic field of liquid mercury.

Strict top-down view of a copper pyramid resting in a shallow pool of opaque white paint.

Strict top-down view of a wet passion fruit resting in a shallow pool of purple oil.

Vertical aerial shot of a stationary waxed lemon.

Bird's-eye view of a shiny green coconut partially submerged in a high-contrast pool of milky juice.

Simulation
PhysGen3D
WonderPlay
MotionCraft
WanMove
Ours (WanMove)

Bird's-eye view showing a gold coin resting calmly on an abstract field of crystal clear milk, creating subtle interference patterns.

Bird's-eye view where a fresh red pear rotates rapidly on top of a turbulent, storm-tossed sea of milky oil, flinging droplets outwards due to centrifugal force.

Direct overhead view of a black ceramic cube resting in a shallow pool of purple liquid.

Flat-lay style shot capturing a waxed coconut smashing into a heavy, moving mass of milky slime.

Flat-lay style shot where a matte rubber thick ring rotates rapidly on top of a stagnant, still pond of translucent teal paint, flinging droplets outwards due to centrifugal force.

Strict overhead view of a bright yellow pomegranate slowly drifting across a perfectly uniform, monochromatic field of liquid mercury.

Strict top-down view of a copper pyramid resting in a shallow pool of opaque white paint.

Strict top-down view of a wet passion fruit resting in a shallow pool of purple oil.

Vertical aerial shot of a stationary waxed lemon.

Bird's-eye view of a shiny green coconut partially submerged in a high-contrast pool of milky juice.

BibTeX

@article{fluidgenzero2026,
  title     = {Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training},
  author    = {Huang, Hong and Liu, Yuqiu and You, Chenyu and Martin, Daniel and Zou, Chuhang and Chen, Wuyang},
  journal   = {arXiv preprint arXiv:2610.10984},
  year      = {2026},
}