README.md

---
license: other
license_name: mixed
license_link: https://huggingface.co/datasets/CInterno/GeoPhys-features#licences
pretty_name: GeoPhys features and videos
tags:
- video
- video-generation
- intuitive-physics
- physical-plausibility
- physics-iq
- representation-geometry
- frozen-encoders
configs:
- config_name: videos
  data_files:
  - split: train
    path:
    - videos/metadata.csv
    - videos/candidates/*/*.mp4
---

# GeoPhys features and videos

The data released with GeoPhys
([paper](https://arxiv.org/abs/2606.20707), [code](https://github.com/ChristianInterno/GeoPhys)):

- **3,168 generated videos** (`videos/`): 16 candidates of [MAGI-1 24B](https://huggingface.co/sand-ai/MAGI-1)
  (video-to-video) for each of the 198 scenarios of
  [Physics-IQ](https://github.com/google-deepmind/physics-IQ-benchmark), each with its Physics-IQ
  score. Any way of choosing among the candidates (a verifier, a reward model, a plausibility
  metric) can be evaluated on Physics-IQ without generating a single video. The same videos
  are also on [Google Drive](https://drive.google.com/drive/folders/1l8gnrSPCD2wxke8SbDU_rnze8wy5kM8B).
- **Frame-trajectory features** (`readout/`, `all_layers/`) of the videos of three physics
  benchmarks, extracted with the four frozen image encoders used by GeoPhys.

## Contents

| folder | what | stores | size |
|---|---|---|---|
| `videos/` | the 3,168 Physics-IQ candidates, with their scores in `videos/metadata.csv` | | 1.0 GB |
| `readout/` | features at the readout layer of each encoder: DINOv2 block 12, DINOv3 block 18, CORnet-S IT, VOneNet V1 block | 32 | 15.3 GB |
| `all_layers/` | features at every layer: 24 blocks for DINOv2 and DINOv3; V1, V2, V4, IT for CORnet-S; the V1 block and ResNet-50 layers 1-4 for VOneNet | 32 | 212.1 GB |

## Videos

`videos/candidates/<scenario_id>/candidate_XXX.mp4` holds candidates 000 to 015 of each
scenario (198 scenarios: 66 scenes, each filmed from three perspectives). Each video: 1280x720 pixels, 24 fps, 120 to 120 frames.

Columns of `videos/metadata.csv`, one row per video:

- `file_name`: path of the video, relative to `videos/`
- `scenario_id`: the Physics-IQ scenario, e.g. `0001_perspective-left_trimmed-ball-and-block-fall`
- `scene`, `perspective`: the two parts of the scenario id
- `candidate`: candidate number, 0 to 15
- `physics_iq_score`: the Physics-IQ score of the candidate, computed for each candidate on its own at 24 fps (higher is better)
- `picked_by_geophys`, `picked_by_geophys_plus`, `picked_by_geophys_plusplus`: whether that GeoPhys configuration picks this candidate for its scenario

The same scores, one row per scenario, are in the code repository as
[`data/magi_v2v_16_per_seed_24fps.csv`](https://github.com/ChristianInterno/GeoPhys/blob/main/data/magi_v2v_16_per_seed_24fps.csv).

Reference points: the mean `physics_iq_score` of one pick per scenario, and the official
Physics-IQ code run on the picked videos.

| pick | mean score of the picks | official Physics-IQ | Physics-IQ Verified |
|---|---|---|---|
| First candidate | 53.25 | – | – |
| Random candidate (expected) | 53.76 | – | – |
| GeoPhys | 63.59 | 62.81 | 58.93 |
| GeoPhys+ | 64.36 | 62.70 | 58.63 |
| GeoPhys++ | 65.51 | 64.58 | 59.90 |
| Best candidate (oracle) | 73.14 | – | – |

## Features

Each feature folder holds one store per `<dataset>_<encoder>_<preprocessing>`:

| dataset | videos |
|---|---|
| `likephys` | [LikePhys](https://github.com/yuanjianhao508/likephys): 800 plausible/implausible pairs, 920 videos |
| `intphys2` | [IntPhys 2](https://huggingface.co/datasets/facebook/IntPhys2), Main set: 506 pairs, 1,012 videos |
| `v2v`, `v2v16` | the 3,168 Physics-IQ candidates in `videos/`; `v2v` holds candidates 0-7, `v2v16` candidates 8-15 |

Encoders: `dinov2` (facebook/dinov2-large), `dinov3` (facebook/dinov3-vitl16-pretrain-lvd1689m),
`cornet` (CORnet-S), `vonenet` (VOneNet, ResNet-50 back end). Preprocessing: `direct` resizes
each frame to 224x224; `hf` resizes the short side to 256 and centre-crops 224x224. Every
video is decoded to at most 60 frames (longer videos are subsampled evenly).

### Files of a store

| file | shape | content |
|---|---|---|
| `gram.npy` | `[N, L, A, 60, 60]` float32 | Gram matrix `G[t, s] = <z_t, z_s>` of each video's frame trajectory, per layer and per frame feature ("atom": `flat` = all patch tokens, `mean` = mean patch token, `cls` = CLS token; the atoms present are listed in `meta.json`) |
| `t2_{raw,l2,spd}_sq.npy` | `[N, L, P, 59]` float32 | squared step of every spatial token between consecutive frames: raw tokens, l2-normalised tokens, or raw steps divided by the token's mean speed |
| `t2_{raw,l2,spd}_dot.npy` | `[N, L, P, 58]` float32 | inner product of consecutive steps of every token |
| `done.npy` | `[N]` bool | extraction completed for each video |
| `meta.json` | | `videos` (paths relative to the benchmark's own video folder), `pairs` (`vi`, `ii`: plausible and implausible video of each pair) or `scenarios` and `index` (scenario and candidate number of each video), `layers`, `atoms`, `P` (tokens per frame) |

N videos, L layers, A atoms, P spatial tokens (196 for the ViTs; 256 for the convolutional
networks, whose feature maps are pooled to 16x16). Videos shorter than 60 frames are padded
with zeros. Every distance, angle and curvature along a trajectory follows from these arrays
without the features themselves; see `geophys/signals.py`.

## Use

```bash
pip install -U huggingface_hub
hf download CInterno/GeoPhys-features --repo-type dataset --include "videos/*" --local-dir geophys_features    # the videos
hf download CInterno/GeoPhys-features --repo-type dataset --include "readout/*" --local-dir geophys_features   # readout features

git clone https://github.com/ChristianInterno/GeoPhys && cd GeoPhys && pip install -e .
python -m geophys.detect --stores ../geophys_features/readout    # detection tables
python -m geophys.verify --stores ../geophys_features/readout    # Physics-IQ table
python -m geophys.select --videos ../geophys_features/videos/candidates --out picks.csv   # GeoPhys on the videos
```

A single store: `--include "all_layers/likephys_dinov3_direct/*"`.

Evaluate your own selector on the videos:

```python
import pandas as pd

m = pd.read_csv("geophys_features/videos/metadata.csv")
m["mine"] = [my_plausibility("geophys_features/videos/" + f) for f in m.file_name]   # higher = better
picks = m.loc[m.groupby("scenario_id")["mine"].idxmax()]
print("Physics-IQ of the picks: %.2f" % picks.physics_iq_score.mean())
```

Read the features:

```python
import numpy as np
from geophys.signals import gram_signals

G = np.load("geophys_features/readout/likephys_dinov3_direct/gram.npy", mmap_mode="r")
g = np.array(G[:, 0, 1], np.float64)        # mean-token trajectories at the readout layer
g[g == 0] = np.nan                           # padding
curvature = gram_signals(g)["curv"]          # one value per video
```

## Licences

The videos and features are derived from third-party videos and models, and each part
follows the terms of its sources:

- `videos/`: generated with MAGI-1 ([arXiv:2505.13211](https://arxiv.org/abs/2505.13211); Sand AI, Apache 2.0) from the input videos of Physics-IQ ([arXiv:2501.09038](https://arxiv.org/abs/2501.09038); Google DeepMind, CC BY 4.0); released under CC BY 4.0.
- `intphys2_*`: derived from IntPhys 2, released under CC BY-NC 4.0. Non-commercial use only, with attribution.
- `v2v_*`, `v2v16_*`: derived from the videos in `videos/`, generated by MAGI-1 (Apache 2.0) from Physics-IQ inputs (CC BY 4.0).
- `likephys_*`: derived from the LikePhys benchmark videos; use them under that benchmark's terms.
- `*_dinov3_*`: produced with DINOv3 and distributed under the DINOv3 License (copy in `LICENSE-DINOv3.md`). Built with DINOv3.

Unless the terms above say otherwise, this dataset is released under CC BY-NC 4.0.

## Citation

```bibtex
@misc{internò2026geophysgeometryphysicalplausibility,
      title={GEOPHYS: The Geometry of Physical Plausibility},
      author={Christian Internò and Alexander Pondaven and Habon Issa and Fabio Pizzati and Francesco Pinto and Markus Olhofer and Ivan Laptev and Philip Torr and Eero P. Simoncelli and Barbara Hammer and David Klindt},
      year={2026},
      eprint={2606.20707},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.20707},
}
```