README.md

metadata

pretty_name: GeoFlip Preview
language:
  - en
license: other
license_name: restricted-research-use
license_link: https://huggingface.co/datasets/DreamingLow/GeoFlip/blob/main/LICENSE.txt
tags:
  - image
  - video
  - physics
  - image-to-video
  - benchmark
  - arxiv:2610.42002
configs:
  - config_name: inputs
    data_files:
      - split: evaluation
        path: first_frames/**

GeoFlip Preview

Benchmarking Geometry-Sensitive Physics in Video Generation
Paper · Code · Project page

482 paired tests · 964 core reference videos · 6 physical categories

Authors: Yuqi Li*, Junming Liu*, Bowen Hu, Ruize Niu, Huiran Duan, Ming-Hsuan Yang, Yingli Tian, Hao Wu†.
*Equal contribution. †Corresponding author.

GeoFlip includes first frames, video-generation prompts, simulation references, evaluation labels, baseline predictions, and scoring code. Alongside the 964 core conditions, 16 neutral controls are provided separately, for 980 conditions in total.

Quick start

Follow the scoring quickstart. To use the toolkit included with this dataset:

python -m pip install -e './toolkit[download,video,http]'

Download scripts support model inputs only, evaluator metadata only, or the complete package, pinning one HF commit and verifying SHA-256 checksums. Give the model only inputs/manifest.jsonl fields and the selected first frames. Never provide evaluation truth or reference futures to a generator.

To score existing generated videos, create a submission JSON mapping each opaque condition id to a native video file. Validate, obtain event/structure judgments, and run the deterministic scorer.

geoflip validate --data . --subset paper-main --submission my-model/submission.json --out validation.json
geoflip score --data . --subset paper-main --predictions my-model/judgments.jsonl --out scores.json
geoflip baselines --data . --subset paper-main --judge gpt-6-astra --out reproduced-main.json

The last command reproduces the event and structure scores of all 16 paper models without API calls. All 32 main-table values match after rounding.

Files

Path	Content
inputs/manifest.jsonl	Model-facing first-frame paths, generation prompts and requested timing
first_frames/	980 input PNG files and image-folder metadata
references/	980 reference MP4 files for evaluation
evaluation/truth.jsonl	Labels, task definitions, observation times, and prompt metadata
evaluation/pairs.jsonl	482 pair memberships and same-first-frame flags
splits/	Frozen condition-id lists
prompts/	Judge template and task/label definitions
baselines/	Baseline judgments and the paper results table
docs/	Quickstart, metrics, frame sampling, and data format
toolkit/	Installable evaluation and download code with tests
SHA256SUMS.json	Integrity hashes for all included files except itself

Scopes

Scope	Conditions	Meaning
core	964	Paired-reference inventory, 482 pairs
auxiliary	16	Neutral controls, separate from the core count
paper-pool	720	Historical selected inference pool, including four controls
paper-main	442	438 core conditions and four controls, across 34 tasks
cross-judge	346	Earlier 342 core conditions and four controls

Use the fixed ID lists in splits/ for comparisons. Original partitions and pair information are recorded in the evaluator metadata.

Metrics and prompts

Event correctness is class-balanced within each task and then task-macro averaged. Structural integrity is the task-macro fraction explicitly judged clear; uncertain, flagged and missing outputs remain in the denominator. See EVALUATION.md.

Generation prompts are in inputs/manifest.jsonl; judge prompts and task definitions are in prompts/. See DATA_FORMAT.md for original prompt text and video-continuation adaptations.

The paper's PSNR/SSIM columns use a separate 40-clip benchmark. Those aggregate numbers are preserved, but that benchmark's reference media are not redistributed here. GeoFlip reference-similarity diagnostics are labeled separately and must not replace the 40-clip values in the paper ranking.

Usage terms

Dataset usage terms are in LICENSE.txt.

The paper's Opus evaluation was run separately; its raw judgments are not included here. Additional historical predictions are identified as Qwen3-VL-32B-Instruct according to their source metadata. These are separate results. The included GPT predictions reproduce all 32 main-table event/structure values.

Citation

@misc{li2026geoflippreview,
  title={GeoFlip Preview: Benchmarking Geometry-Sensitive Physics in Video Generation},
  author={Yuqi Li and Junming Liu and Bowen Hu and Ruize Niu and Huiran Duan
          and Ming-Hsuan Yang and Yingli Tian and Hao Wu},
  year={2026},
  eprint={2610.42002},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}