PPAVG Read paper

PHYSICS-GROUNDED VIDEO GENERATION · 2026

Phase-Aware Video Generation
for Physics-Grounded
Dynamics and Interactions

Different phases follow different motion laws.
We learn their dynamics and interactions as 3D trajectories to guide video generation.

Jingfeng Ou1,2Kun Wang3Rui Zhao3Jingwei Guan4Limin Wang1Chao Dong2Xingyu Zeng2,*

1 Nanjing University2 Shenzhen University of Advanced Technology3 SenseTime4 Shenzhen Technology University

* Corresponding author

Paper Explore videos Code Model weights

HOW IT WORKS

Predict physical trajectories.
Use them to guide video.

PAVG follows the learned-trajectory route to video generation: physics conditions determine motion in an explicit 3D representation, and a video model supplies the visual appearance. The key design is to preserve phase-specific dynamics while learning their interaction.

  1. 01 / INPUTImage + physical conditions

    Lift the initial scene into phase-labeled 3D points.

  2. 02 / DYNAMICSPhase-aware prediction

    Solid and gas branches, connected by solid-to-gas attention.

  3. 03 / INTERFACEShared 3D trajectories

    Project the predicted point motion into image space.

  4. 04 / GENERATIONTrajectory-guided video

    Condition the video generator on the projected motion.

PAVG overview: physical scene initialization, phase-aware trajectory prediction, and trajectory-guided video generation.
Overview from the paper. Click the figure to view it at full size.
Explore the trajectory architectureSolid and gas trajectory branches with spatiotemporal attention and a solid-to-gas cross-phase pathway.

Phase-specific pretraining is followed by coupled training. The learned representation keeps fixed point identities across future frames. In the present model, cross-phase influence runs from solid to gas.

THE IDEA

Respect the differences.
Model the interactions.

Solids may retain structure, deform, or respond to contact. Gases evolve through transport, pressure, dissipation, and buoyancy. A useful physical motion model must preserve these differences while capturing how one phase influences another.

01

Phase-specific dynamics

Separate solid and gas predictors learn distinct motion patterns, instead of requiring one undifferentiated dynamics model to explain both.

02

Cross-phase interactions

Spatiotemporal cross-attention conveys solid motion to the gas branch, modeling where, when, and how the surrounding gas responds.

03

A shared trajectory interface

Both branches predict 3D point trajectories. Their projections provide motion guidance to a video generator while retaining an explicit physical intermediate representation.

Our experiments study solid dynamics, gas dynamics, and solid-to-gas interactions. Here, “phase” means a modeled state of matter; it does not mean a phase transition.

SEE THE MOTION

From deformation to interaction.

Selected examples with their physical inputs and predicted motion.

Pot lid and steam

Solid–gas interaction

Initial conditions
Projected solid force and/or gas initial velocity for Pot lid and steam
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Cushion and haze

Solid–gas interaction

Initial conditions
Projected solid force and/or gas initial velocity for Cushion and haze
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Soccer ball and dust

Solid–gas interaction

Initial conditions
Projected solid force and/or gas initial velocity for Soccer ball and dust
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Fighter jet and clouds

Solid–gas interaction

Initial conditions
Projected solid force and/or gas initial velocity for Fighter jet and clouds
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Smoke: up-left

Gas dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Smoke: up-left
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Smoke: down-right

Gas dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Smoke: down-right
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Mist: up-right

Gas dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Mist: up-right
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Mist: up-left

Gas dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Mist: up-left
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Sand bear

Solid dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Sand bear
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Chair

Solid dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Chair
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Giraffe

Solid dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Giraffe
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Penguin

Solid dynamics

Initial conditions
Projected solid force and/or gas initial velocity for Penguin
Predicted trajectories
Generated video

Cyan: solid force / points. Gold: gas initial velocity / points. Arrow tails mark the solid force point or the gas centroid. Arrow lengths illustrate direction.

Original videoTrajectory videoAnnotated initial image

Explore 12 examples, from initial conditions to predicted 3D trajectories to generated video. The initial image marks solid force and uniform gas initial velocity. Both videos play together. Example details.

DATA & EVALUATION

What the dataset contains.

Initial 3D states, physical conditions and future point trajectories cover solid motion, gas transport and their interaction.

700K+simulated training trajectories
700held-out trajectory sequences
32video-generation scenes

Solid points are orange. Gas points are blue.

Solid: an elastic ball

Dot: force position. Arrow: force direction.
|f| = 3.03 per particle in simulation units.

Elastic materials, plasticine, sand and rigid bodies under force or gravity. Here, a local force acts on 131 selected points and gravity is disabled.

Force condition

f = (1.451, −1.008, 2.462) per selected particle. Initial application center: (5.280, 4.315, 4.331). The dot follows that material point. The red patch marks the selected region. Arrow length is illustrative.

500 trajectory tests · 8 video scenes

Gas transport

Arrow: initial flow at the cloud center.
|v₀| = 2.16 in simulation velocity units.

Gas evolution with initial-velocity conditions, generated using MantaFlow. Fixed point identities retain the motion of individual markers over time. The arrow shows the cloud-wide initial velocity condition, with the centroid used as a visual reference.

100 trajectory tests · 8 video scenes

Interaction: small rigid ball

Blue: gas force. Purple: ball force.
Gas: 10 pN right · Ball: 7 pN left

The gas cloud and ball start apart and move toward each other. Gas paths bend around the moving ball.

Force condition

Both forces remain constant over 0–4.8 s. Gas force: (10, 0, 0) pN in a fixed spherical region centered at (−0.40, 0, 0) mm with radius 0.08 mm. Solid force: (−7, 0, 0) pN at the moving ball’s center. The ball starts at the origin with radius 0.10 mm. All 2048 gas markers start in one 3D ellipsoid centered at (−0.35, 0, 0) mm, with radii (0.13, 0.165, 0.165) mm. The initial gas-to-surface gap is at least 0.126 mm. No additional gas markers start near the ball. The ball’s velocity follows force balance in the carrier gas flow. Moving no-slip conditions change gas trajectories around the ball. This is quasi-steady creeping flow with negligible inertia, and the displayed gas points mark a parcel within a continuous carrier. The camera remains fixed, the solid uses only points, and the two arrows retain the initial external-force directions and application positions. Arrow lengths are illustrative. The 49 saved states loop over 6.125 s of playback. This supplementary simulation is separate from the frozen dataset and reported evaluations.

100 trajectory tests · 16 video scenes

The solid–gas video is an additional low-Re Stokes-flow illustration, outside the frozen dataset and reported evaluations. Trajectory tests use fixed point identities and common normalization. The 32 video-generation scenes form a separate evaluation setting.

RESULTS THAT CAN BE MEASURED

Quantitative experiments.

Complete trajectory, video-generation and ablation results from the paper.

Trajectory prediction

All ten predictors and all applicable settings. vIoU measures occupied-volume overlap; CD measures Chamfer distance; Corr measures correspondence-aware pointwise L2 error. Higher vIoU and lower CD and Corr are better.

Trajectory prediction on the fixed 700-sequence benchmark.
MethodOverallSingle-phase solidSingle-phase gasSolid–gas interaction
vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓
SFBC——————0.4838740.1209060.274170———
DLF——————0.4835050.0536500.241004———
SEGNN——————0.5618930.0070600.176081———
Motion2VecSets———0.1411720.8429130.656012——————
MDM———0.3573440.0977700.246086——————
PhysCtrl———0.5883720.0228820.080403——————
GNS0.4652410.0274800.2098790.4224630.0397140.1633460.5596940.0071630.1762190.4394040.0315210.249974
HGATSolver0.4492270.0680790.2233030.3066640.1639930.3095130.5704280.0062770.1169320.4599070.0510230.233384
Fisale0.5181430.0362050.1824530.4787800.0544900.1447280.5788050.0058740.1176750.5074920.0422290.233703
PAVG0.6963430.0066300.0624330.6119640.0172940.0683610.7933520.0013700.0615150.6900280.0039280.059928

Overall = 25% solid + 25% gas + 50% interaction. Interaction metrics use concatenated solid and gas point sets. — means the method does not apply to that setting. Values retain the paper’s six decimal places. Bold marks the best reported result.

Video generation scores

Five multimodal judges score semantic adherence (SA), physical commonsense (PC) and video quality (VQ). Scores average equally over 32 scenes and five judges.

All eleven methods. Scores range from 1 to 5; higher is better.
MethodSA ↑PC ↑VQ ↑
HunyuanVideo-1.53.113.033.66
Wan2.2-I2V-A14B3.053.083.53
CogVideoX1.5-5B-I2V2.262.462.93
DragAnything2.402.452.69
ObjCtrl-2.5D2.282.272.83
FlashMotion3.173.183.58
Wan-Move2.091.992.25
Force Prompting2.032.242.86
StreamForce1.912.042.78
PhysCtrl3.102.983.33
PAVG3.393.283.61

Bold marks the best result; underlining marks the second best.

Results from each model judge

Each judge evaluates the same eleven methods on the same 32 scenes using the shared rubric.

All five judges and all three criteria; higher is better.
MethodGPT-4oGPT-5.5GPT-5.6-solGemini-3.8-flashDeepSeek-v4.1-flash
SA ↑PC ↑VQ ↑SA ↑PC ↑VQ ↑SA ↑PC ↑VQ ↑SA ↑PC ↑VQ ↑SA ↑PC ↑VQ ↑
HunyuanVideo-1.53.633.633.883.062.913.592.942.753.693.003.003.632.942.883.53
Wan2.2-I2V-A14B3.313.503.813.163.193.442.532.413.342.943.003.533.313.283.53
CogVideoX1.5-5B-I2V2.752.943.412.132.312.721.842.192.752.312.472.912.252.412.84
DragAnything2.782.813.062.382.312.501.942.092.412.342.442.722.562.592.78
ObjCtrl-2.5D2.592.663.032.092.032.752.162.062.752.502.532.882.032.062.72
FlashMotion3.283.443.753.533.413.723.002.883.592.943.033.473.093.133.38
Wan-Move2.472.473.002.031.841.811.751.631.782.342.342.501.881.692.16
Force Prompting2.282.412.882.092.253.001.692.032.722.132.343.031.972.162.69
StreamForce2.342.412.971.751.812.751.441.722.502.002.192.942.032.092.72
PhysCtrl3.413.473.782.882.813.133.222.843.343.193.133.312.812.633.09
PAVG3.753.784.003.413.253.663.443.193.533.343.223.533.002.943.31

Human preference study

Participants select the best videos for physical plausibility and overall video quality. A selection rate counts every tied choice, so rates across methods may sum to more than 100%.

15 participants, 32 scenes, 480 responses per criterion. Tied best choices are allowed.
MethodPhysical plausibilityOverall video quality
HunyuanVideo-1.523.12%21.67%
Wan2.2-I2V-A14B18.12%21.04%
CogVideoX1.5-5B-I2V3.54%5.00%
DragAnything5.62%3.96%
ObjCtrl-2.5D3.96%4.17%
FlashMotion13.75%12.50%
Wan-Move5.42%3.54%
Force Prompting0.42%0.83%
StreamForce2.92%2.71%
PhysCtrl15.00%14.17%
PAVG32.50%32.08%

Video scores and human preferences describe the paper’s fixed evaluation release. They were not recomputed for the selected demonstrations above.

Phase-specific modeling

Separate solid and gas predictors are compared with a unified predictor at matched cumulative training exposure per phase.

Solid trajectories: all three predictors.
PredictorvIoU ↑CD ↓Corr ↓
Solid-only0.5883720.0228820.080403
Gas-onlyN/AN/AN/A
Unified0.5711500.0248510.086428
Gas trajectories: all three predictors.
PredictorvIoU ↑CD ↓Corr ↓
Solid-onlyN/AN/AN/A
Gas-only0.7992630.0012810.059713
Unified0.7224830.0019840.083933

Interaction pathway

The same PAVG checkpoint is evaluated with solid-to-gas cross-attention enabled or disabled. Solid predictions remain identical.

Interaction-pathway ablation: joint output and both individual phases.
VariantJoint solid–gasInteraction: solidInteraction: gas
vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓
PAVG0.6900280.0039280.0599280.7280560.0115930.0497340.6422960.0035370.070121
w/o solid-to-gas Cross-Attention0.5407550.0160720.1788300.7280560.0115930.0497340.4284350.0671360.307926
Temporal trajectory comparison of ground truth, PAVG without solid-to-gas cross-attention and full PAVG on three held-out interaction cases
Fading point clouds and short trails show motion over time on three held-out interaction cases.

Stage 2 initialization

The same dual-branch architecture and reduced-loss training recipe use either random or phase-specific initialization. This controlled ablation has a different loss recipe from the main model.

Stage 2 initialization ablation: every setting and metric.
InitializationOverallSingle-phase solidSingle-phase gasSolid–gas interaction
vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓vIoU ↑CD ↓Corr ↓
Random0.6449720.0081630.0827040.5822680.0221990.0809270.7102860.0020360.0963000.6436670.0042090.076795
Phase-specific0.6909640.0065860.0631350.6179500.0167840.0679020.7834330.0014440.0626110.6812360.0040590.061014

Interaction physical constraint

The coupling objective is enabled or removed under the matched reduced-loss recipe.

Coupling-loss ablation on solid–gas interaction.
ObjectivevIoU ↑CD ↓Corr ↓
Without coupling loss0.6776080.0039480.061509
With coupling loss0.6812360.0040590.061014

Source: PAVG paper and experiment appendix. Structured quantitative tables.

CURRENT SCOPE

A step toward richer
physical worlds.

The current video pipeline initializes a 2.5D scene proxy from an image. Our experiments validate solids, gases, and one-way solid-to-gas influence; they do not establish full bidirectional coupling or general multiphase physics.

Liquids, phase transitions, stronger scene reconstruction, long-horizon prediction, and tighter preservation of trajectories in the generated pixels remain future directions.

Perceptual plausibility, human preferences, and numerical trajectory agreement provide complementary evidence. None alone certifies physical correctness in a real scene.

EXPLORE & BUILD

Resources.

Read the paper, explore the code, and download pretrained models.

Paper

Full method, experiments, and appendix.

arXiv:2610.11791

Benchmark protocol

Trajectory settings, video evaluation, metrics, and baseline details.

Available in the paper

Code

Installation, example inputs, video generation, training and evaluation.

View on GitHub ↗

Model weights

Pretrained joint trajectory predictor, solid and gas specialists, and matching configs.

Download on Hugging Face ↗

PAVG source code and trajectory checkpoints are released under Apache-2.0. Third-party components retain their respective terms.

Citation

@misc{ou2026pavg,
  title={Phase-aware video generation for physics-grounded dynamics and interactions},
  author={Jingfeng Ou and Kun Wang and Rui Zhao and Jingwei Guan and
          Limin Wang and Chao Dong and Xingyu Zeng},
  year={2026},
  eprint={2610.11791},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.11791}
}