File size: 2,666 Bytes
2af8cc6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---

license: cc-by-nc-sa-4.0
language:
- en
pretty_name: GimbalDiffusion SpatialVID-Extreme
size_categories:
- n<1K
task_categories:
- text-to-video
tags:
- video
- camera-control
- camera-pose
- benchmark
- spatialvid
---


# SpatialVID-Extreme

The 138-sample extreme-camera-motion benchmark for **GimbalDiffusion: Gravity-Aware Camera Control for Video Generation**.

- [Project page](https://lvsn.github.io/GimbalDiffusion/)
- [Paper](https://arxiv.org/abs/2512.09112)
- [SpatialVID-HQ source dataset](https://huggingface.co/datasets/FelixYuan/SpatialVID-HQ)

## Contents

- `spatialvid_extreme.tar`: the complete benchmark in one archive, unpacking directly into the layout below.
- `test_index.json` and `test_samples/`: the canonical 138 prompts, seeds, absolute/relative camera matrices, intrinsics, and source identifiers.
- `assets/ground_truth_raw/` and `assets/ground_truth_for_fid_fvd/`: reference videos used by the evaluation.

Generated method outputs, legacy indices, derived hover/trajectory videos, and metric files are intentionally excluded.

## Download and evaluate

```bash

hf download lefreud/GimbalDiffusion-SpatialVID-Extreme spatialvid_extreme.tar \

  --repo-type dataset \

  --local-dir data/spatialvid_extreme

tar -xf data/spatialvid_extreme/spatialvid_extreme.tar \

  -C data/spatialvid_extreme

python -m src.scripts.evaluate_spatialvid_extreme \

  --data-root data/spatialvid_extreme \

  --method ours_ucpe \

  --negative-prompt ours_look_down

```

Complete the matching environment and checkpoint setup in the [GimbalDiffusion README](https://github.com/lvsn/GimbalDiffusion#readme) before generating outputs. Each sample JSON contains `semantic_prompt`, `seed`, `c2w_absolute`, `c2w_relative`, and `intrinsics`; evaluation selects the pose convention required by the chosen method.

## Limitations and license

This is a deliberately difficult, small diagnostic benchmark rather than a representative sample of all prompts or camera motions. Scores should be interpreted together with the protocol in the paper and supplement.

The benchmark is derived from SpatialVID-HQ material distributed under CC-BY-NC-SA-4.0 and is released under the same license. See [LICENSE.md](LICENSE.md).

```bibtex

@inproceedings{fortierchouinard2026gimbaldiffusion,

  title = {GimbalDiffusion: Gravity-Aware Camera Control for Video Generation},

  author = {Fortier-Chouinard, Fr{\'e}d{\'e}ric and Hold-Geoffroy, Yannick and Deschaintre, Valentin and Gadelha, Matheus and Lalonde, Jean-Fran{\c{c}}ois},

  booktitle = {European Conference on Computer Vision (ECCV)},

  year = {2026}

}

```