From a single input video, we reconstruct the fluid and show novel view synthesis results below. We show two novel views for each dataset example.
Future Prediction
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 60 frames (2 seconds) are reconstruction, the last 60 frames (2 seconds) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 30 frames (1 second) are reconstruction, the last 30 frames (1 second) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 30 frames (1 second) are reconstruction, the last 30 frames (1 second) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 30 frames (1 second) are reconstruction, the last 30 frames (1 second) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 30 frames (1 second) are reconstruction, the last 30 frames (1 second) are future
prediction.
Our physics-integrated fluid representation directly allows future prediction by physics simulation on the reconstructed fluid particles. The first 30 frames (1 second) are reconstruction, the last 30 frames (1 second) are future
prediction.
Counterfactual Interaction Simulation
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 50 frames (1.67 seconds) are the reconstruction, the last 56 frames (1.87 seconds) include the added wind effect.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 50 frames (1.67 seconds) are the reconstruction, the last 56 frames (1.87 seconds) include the added wind effect.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 50 frames (1.67 seconds) are the reconstruction, the last 56 frames (1.87 seconds) include the added wind effect.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 50 frames (1.67 seconds) are the reconstruction, the last 56 frames (1.87 seconds) include the added wind effect.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 50 frames (1.67 seconds) are the reconstruction, the last 56 frames (1.87 seconds) include the added wind effect.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 11 frames (0.37 seconds) are the reconstruction, the last 79 frames (2.63 seconds) include the added object.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 11 frames (0.37 seconds) are the reconstruction, the last 79 frames (2.63 seconds) include the added object.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 11 frames (0.37 seconds) are the reconstruction, the last 79 frames (2.63 seconds) include the added object.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 11 frames (0.37 seconds) are the reconstruction, the last 79 frames (2.63 seconds) include the added object.
Our physics-integrated fluid representation directly allows counterfactual interaction simulation by physics simulation on the reconstructed fluid particles. We enable this new task which was not possible with prior methods. The first 11 frames (0.37 seconds) are the reconstruction, the last 79 frames (2.63 seconds) include the added object.
Re-simulation on Novel Views
(Click to show/hide)
We use reconstructed velocity to advect the reconstructed appearance.
Re-simulation on In-the-wild Scene
(Click to show/hide)
We use reconstructed velocity to advect the reconstructed appearance.
Ablation Study on Novel-View Video Synthesizer
(Click to show/hide)
NVS: novel-view video synthesizer, GVR: generative video refinement, LVG: long video generation
Ablation Study on Physics Losses
(Click to show/hide)
Abstract
We introduce FluidNexus, a framework for reconstructing and predicting 3D fluid dynamics
from a single video by synthesizing novel-view videos for reference, integrating video
generation with physics simulation. FluidNexus combines realistic video synthesis via
diffusion-based refinement and a physics-integrated particle representation for fluid
reconstruction and prediction. We validate our approach with two new real-world fluid datasets
and enable dynamic novel view synthesis, future prediction, and interaction simulation. Code and
datasets will be released.
Method Overview
FluidNexus in reconstruction. We represent 3D fluid dynamics by a novel physics-integrated representation, two-layer fluid particles. During reconstruction, from a single video, we synthesize multiple novel-view videos as references for 3D fluid reconstruction and then optimize the two-layer particle fluid representations over time. The optimization is supervised using the multi-view video frames to compute the visual loss and the physics constraints to compute the physics loss. Our reconstruction output is the 3D fluid appearance and velocity fields over all input frames.
The FluidNexus-Smoke and FluidNexus-Ball Datasets
Our FluidNexus-Smoke and FluidNexus-Ball datasets include 120 scenes for each. Each
scene has 5 synchronized multi-view videos where the cameras are placed along a
horizontal arc of approximately 120°.
BibTex
@InProceedings{gao2025fluidnexus,
title = {FluidNexus: 3D Fluid Reconstruction and Prediction from a Single Video},
author = {Gao, Yue and Yu, Hong-Xing and Zhu, Bo and Wu, Jiajun},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2025},
}
