Datasets:

hhkka
/

FluidGen-Benchmark

---
pretty_name: FluidGen Benchmark
task_categories:
- image-to-video
tags:
- physics
- fluid-simulation
- video-generation
- benchmark
size_categories:
- n<1K
---
<h1 align="center">🌊 FluidGen Benchmark</h1>

<h3 align="center">Fluid-object interaction scenes with full physics simulation rollouts</h3>

<p align="center">
  <a href="https://arxiv.org/abs/2610.10984"><img src="https://img.shields.io/badge/arXiv-2610.10984-b31b1b?logo=arxiv&logoColor=white" alt="arXiv"></a>
  <a href="https://hongh0.github.io/Fluid-Gen-Zero/"><img src="https://img.shields.io/badge/Project-Page-1f72b8?logo=googlechrome&logoColor=white" alt="Project Page"></a>
  <a href="https://github.com/hongh0/Fluid-Gen-Zero"><img src="https://img.shields.io/badge/GitHub-Code-181717?logo=github&logoColor=white" alt="Code"></a>
</p>

<p align="center"><img src="assets/samples.jpg" width="100%" alt="FluidGen Benchmark scenes"></p>

The benchmark introduced in **Fluid-Gen-Zero: Grounding Pretrained Video
Generators in Physics without Training**. Each of its **43 scenes** pairs an
input image and a prompt with a physics simulation rollout of that scene, so
generated videos can be measured against where the object and the fluid
*should* go, not only against how plausible they look.

| | |
|---|---|
| 🎬 Scenes | 43, across 10 primary fluid categories |
| ⏱️ Rollout | 401 frames per scene (0.01 s of simulated time per frame) |
| 🧊 3D points | object and fluid particle trajectories, every frame |
| 🎭 Masks | projected object and fluid masks, every frame |
| 🌀 Optical flow | projected simulation optical flow, every frame |
| 🏷️ Annotations | fluid category, object material, interaction intensity, camera view, background, tags |

## 🚀 Quick start

```python
import json, numpy as np
from huggingface_hub import snapshot_download
root = snapshot_download("hhkka/FluidGen-Benchmark", repo_type="dataset")
samples = json.load(open(f"{root}/benchmark.json"))
s = samples[0]
print(s["prompt"], s["primary_fluid_category"])
obj = np.load(f"{root}/{s['simulation_dir']}/fg_traj.npz")["xyz"]      # (401, N_obj, 3)
fluid = np.load(f"{root}/{s['simulation_dir']}/fluid_traj.npz")["xyz"]  # (401, N_fluid, 3)
```

To generate videos on the benchmark with Fluid-Gen-Zero, see the
[code repository](https://github.com/hongh0/Fluid-Gen-Zero).

## 📁 Layout

```
benchmark.json                      one entry per scene (below)
samples/<sample_id>/
  image.png                         input image I_0 (= first frame)
  simulation/                       simulation rollout, in the layout the code reads as a "bundle"
    fg_traj.npz                     xyz (401, N_obj, 3)    object particles, world frame (z up, metres)
    fluid_traj.npz                  xyz (401, N_fluid, 3)  fluid particles,  world frame
    segmap/fg_mask.mp4              projected object mask, 401 frames, image resolution
    segmap/fluid_mask.mp4           projected fluid mask, 401 frames
    segmap/segmap.mp4               both masks in one colour-coded video
    optical_flow.mp4                projected simulation optical flow (Middlebury colours), 400 frames
    trajectory_2d.npz               object_xy, fluid_xy (401, 2): projected centroids, pixels
    object_trajectory.txt           same object centroid track as "x,y" lines
    fluid_trajectory.txt            same fluid centroid track
    simulation.mp4                  simulation preview (point rendering)
    warped_reference/motion_signal.mp4   appearance-preserving warped reference video (480x832)
    warped_reference/mask.mp4            its object mask
    warped_reference/{first_frame,background,mask0}.png, disp_vis.mp4   intermediate products
    rollout.json                    simulator settings: particles, substeps, domain,
                                    external-force schedule
  scene/
    scene.json                      camera (K, T_world_cam, width, height), object pose / scale,
                                    materials, simulation domain, metric scale
    object.glb                      reconstructed object mesh (SAM 3D Objects), object frame
    object_material.json, fluid_material.json
```

<details>
<summary><b><code>benchmark.json</code> entry</b></summary>

```json
{
  "sample_id": "000_birdseye_view_showing_a_gold_coin_resting_calmly",
  "prompt": "Bird's-eye view showing a gold coin resting calmly on ...",
  "image": "samples/000_.../image.png",
  "simulation_dir": "samples/000_.../simulation",
  "scene_json": "samples/000_.../scene/scene.json",
  "num_frames": 401,
  "primary_fluid_category": "K. MULTI_PHASE_MIXING_SWIRL",
  "annotations": {"object_type": "METAL", "interaction_intensity": "WEAK",
                  "camera_view": "aerial_top_down", "background_style": "minimalist_studio",
                  "secondary_tags": ["coin", "..."], "object_shape_hint": "coin/disk", "...": "..."}
}
```

</details>

## 🌀 Dense optical flow

The dense flow (`(400, 2, H, W)` float32 pixel displacements, about 3 GB per
scene) is not stored. It is a deterministic projection of the particle
trajectories through the camera in `scene.json`, and is re-rendered with the
[Fluid-Gen-Zero code](https://github.com/hongh0/Fluid-Gen-Zero), byte-identical
to the flow used in the paper:

```bash
python fluidgen/simulation/render_conditions.py \
    --scene   samples/<id>/scene/scene.json \
    --sim_dir samples/<id>/simulation \
    --out_dir <work>/<id>/s9_cond          # -> optical_flow.npy
```

## 📊 Statistics

<p align="center"><img src="assets/benchmark_statistics.png" width="100%" alt="Benchmark statistics"></p>

| Primary fluid category | Scenes |
|---|:---:|
| High-viscosity fluid | 10 |
| Paint / ink / dye | 9 |
| Splash / impact / collision | 8 |
| Water stream / downpour | 5 |
| Multi-phase mixing / swirl | 3 |
| Pool / tank / container water | 3 |
| Ocean / sea waves | 2 |
| Rain / mist / spray droplets | 1 |
| River / stream flow | 1 |
| Other fluid | 1 |

Object material: organic 23, metal 7, plastic 5, ceramic 4, wood / stone / fabric / unknown 1 each.
Interaction intensity: weak 10, medium 23, strong 10.
Camera view: aerial top-down 28, oblique 8, close-up macro 4, frontal 2, side 1.

## 📝 Citation

```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},
}
```

	---
	pretty_name: FluidGen Benchmark
	task_categories:
	- image-to-video
	tags:
	- physics
	- fluid-simulation
	- video-generation
	- benchmark
	size_categories:
	- n<1K
	---

	<h1 align="center">🌊 FluidGen Benchmark</h1>

	<h3 align="center">Fluid-object interaction scenes with full physics simulation rollouts</h3>

	<p align="center">
	<a href="https://arxiv.org/abs/2610.10984"><img src="https://img.shields.io/badge/arXiv-2610.10984-b31b1b?logo=arxiv&logoColor=white" alt="arXiv"></a>
	<a href="https://hongh0.github.io/Fluid-Gen-Zero/"><img src="https://img.shields.io/badge/Project-Page-1f72b8?logo=googlechrome&logoColor=white" alt="Project Page"></a>
	<a href="https://github.com/hongh0/Fluid-Gen-Zero"><img src="https://img.shields.io/badge/GitHub-Code-181717?logo=github&logoColor=white" alt="Code"></a>
	</p>

	<p align="center"><img src="assets/samples.jpg" width="100%" alt="FluidGen Benchmark scenes"></p>

	The benchmark introduced in **Fluid-Gen-Zero: Grounding Pretrained Video
	Generators in Physics without Training. Each of its 43 scenes** pairs an
	input image and a prompt with a physics simulation rollout of that scene, so
	generated videos can be measured against where the object and the fluid
	should go, not only against how plausible they look.

	\| \| \|
	\|---\|---\|
	\| 🎬 Scenes \| 43, across 10 primary fluid categories \|
	\| ⏱️ Rollout \| 401 frames per scene (0.01 s of simulated time per frame) \|
	\| 🧊 3D points \| object and fluid particle trajectories, every frame \|
	\| 🎭 Masks \| projected object and fluid masks, every frame \|
	\| 🌀 Optical flow \| projected simulation optical flow, every frame \|
	\| 🏷️ Annotations \| fluid category, object material, interaction intensity, camera view, background, tags \|

	## 🚀 Quick start

	```python
	import json, numpy as np
	from huggingface_hub import snapshot_download

	root = snapshot_download("hhkka/FluidGen-Benchmark", repo_type="dataset")
	samples = json.load(open(f"{root}/benchmark.json"))

	s = samples[0]
	print(s["prompt"], s["primary_fluid_category"])
	obj = np.load(f"{root}/{s['simulation_dir']}/fg_traj.npz")["xyz"] # (401, N_obj, 3)
	fluid = np.load(f"{root}/{s['simulation_dir']}/fluid_traj.npz")["xyz"] # (401, N_fluid, 3)
	```

	To generate videos on the benchmark with Fluid-Gen-Zero, see the
	[code repository](https://github.com/hongh0/Fluid-Gen-Zero).

	## 📁 Layout

	```
	benchmark.json one entry per scene (below)
	samples/<sample_id>/
	image.png input image I_0 (= first frame)
	simulation/ simulation rollout, in the layout the code reads as a "bundle"
	fg_traj.npz xyz (401, N_obj, 3) object particles, world frame (z up, metres)
	fluid_traj.npz xyz (401, N_fluid, 3) fluid particles, world frame
	segmap/fg_mask.mp4 projected object mask, 401 frames, image resolution
	segmap/fluid_mask.mp4 projected fluid mask, 401 frames
	segmap/segmap.mp4 both masks in one colour-coded video
	optical_flow.mp4 projected simulation optical flow (Middlebury colours), 400 frames
	trajectory_2d.npz object_xy, fluid_xy (401, 2): projected centroids, pixels
	object_trajectory.txt same object centroid track as "x,y" lines
	fluid_trajectory.txt same fluid centroid track
	simulation.mp4 simulation preview (point rendering)
	warped_reference/motion_signal.mp4 appearance-preserving warped reference video (480x832)
	warped_reference/mask.mp4 its object mask
	warped_reference/{first_frame,background,mask0}.png, disp_vis.mp4 intermediate products
	rollout.json simulator settings: particles, substeps, domain,
	external-force schedule
	scene/
	scene.json camera (K, T_world_cam, width, height), object pose / scale,
	materials, simulation domain, metric scale
	object.glb reconstructed object mesh (SAM 3D Objects), object frame
	object_material.json, fluid_material.json
	```

	<details>
	<summary><b><code>benchmark.json</code> entry</b></summary>

	```json
	{
	"sample_id": "000_birdseye_view_showing_a_gold_coin_resting_calmly",
	"prompt": "Bird's-eye view showing a gold coin resting calmly on ...",
	"image": "samples/000_.../image.png",
	"simulation_dir": "samples/000_.../simulation",
	"scene_json": "samples/000_.../scene/scene.json",
	"num_frames": 401,
	"primary_fluid_category": "K. MULTI_PHASE_MIXING_SWIRL",
	"annotations": {"object_type": "METAL", "interaction_intensity": "WEAK",
	"camera_view": "aerial_top_down", "background_style": "minimalist_studio",
	"secondary_tags": ["coin", "..."], "object_shape_hint": "coin/disk", "...": "..."}
	}
	```

	</details>

	## 🌀 Dense optical flow

	The dense flow (`(400, 2, H, W)` float32 pixel displacements, about 3 GB per
	scene) is not stored. It is a deterministic projection of the particle
	trajectories through the camera in `scene.json`, and is re-rendered with the
	[Fluid-Gen-Zero code](https://github.com/hongh0/Fluid-Gen-Zero), byte-identical
	to the flow used in the paper:

	```bash
	python fluidgen/simulation/render_conditions.py \
	--scene samples/<id>/scene/scene.json \
	--sim_dir samples/<id>/simulation \
	--out_dir <work>/<id>/s9_cond # -> optical_flow.npy
	```

	## 📊 Statistics

	<p align="center"><img src="assets/benchmark_statistics.png" width="100%" alt="Benchmark statistics"></p>

	\| Primary fluid category \| Scenes \|
	\|---\|:---:\|
	\| High-viscosity fluid \| 10 \|
	\| Paint / ink / dye \| 9 \|
	\| Splash / impact / collision \| 8 \|
	\| Water stream / downpour \| 5 \|
	\| Multi-phase mixing / swirl \| 3 \|
	\| Pool / tank / container water \| 3 \|
	\| Ocean / sea waves \| 2 \|
	\| Rain / mist / spray droplets \| 1 \|
	\| River / stream flow \| 1 \|
	\| Other fluid \| 1 \|

	Object material: organic 23, metal 7, plastic 5, ceramic 4, wood / stone / fabric / unknown 1 each.
	Interaction intensity: weak 10, medium 23, strong 10.
	Camera view: aerial top-down 28, oblique 8, close-up macro 4, frontal 2, side 1.

	## 📝 Citation

	```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},
	}
	```