This repository contains no video or frames — only annotations, questions, build
recipes and model outputs. Source videos must be obtained from ACaM, SpatialVID and
Blender under their own terms. Manual review; non-commercial research only.
The contents listing, question design and known issues ship with the files.

Long-form Camera Motion Bench

English · 简体中文 · 繁體中文

A benchmark for camera-motion understanding over long video — minutes, not seconds.

Almost everything in this area is evaluated on a single short shot. This set moves the question to the long-form setting: what the camera does across a whole clip, when it changes, and in what order. It ships the human-labelled annotations, the question sets, the exact build recipes, and the raw outputs of every model that was run on it, so the evaluation can be reproduced rather than taken on trust.

No video or frames are included. Every stitched video can be rebuilt exactly from the source clips listed in the recipes.

Contents listing, question design and known issues ship with the files and become available once your access request is approved; the accompanying work is not yet published.

Companion models (also gated): Zeteng/Qwen3-VL-8B-long-camera-motion.

Access

Click Request access on this page and complete the form. Requests are reviewed individually.

Sources, licences and credit

Nothing here is video. What is redistributed is annotations, questions, recipes, and the raw outputs of other people's models. Please cite the upstreams you rely on.

Component	Source	Terms
Benchmark labels and recipes	ACaM — Natural Language Camera Movement Understanding, arXiv 2607.03043	ACaM's terms; videos must be obtained from ACaM
Pose-derived timelines	`SpatialVID/SpatialVID-RAW`	SpatialVID's licence (non-commercial research); YouTube videos not redistributed
Open-film pre-labels	Sintel (CC BY 3.0), Tears of Steel (CC BY 3.0), Cosmos Laundromat (CC BY 4.0) — © Blender Foundation, www.blender.org	the films' CC BY licences; download from Blender's official sources
Questions, recipes, statistics	this work	non-commercial research

Models whose raw outputs are redistributed under outputs/:

Model	Licence
`Qwen/Qwen3-VL-8B-Instruct`	Apache-2.0
`Qwen/Qwen3-VL-32B-Instruct`	Apache-2.0
`OpenGVLab/InternVL3_5-8B`	Apache-2.0
`Yuwen2024/Qwen3-VL-8B-camera-motion`	⚠️ no licence stated upstream
`ddz16/CamSFT-8B`	Apache-2.0
`ddz16/CamDistill-8B`	Apache-2.0

@misc{acam,
  title         = {Natural Language Camera Movement Understanding},
  year          = {2026},
  eprint        = {2607.03043},
  archivePrefix = {arXiv}
}

@misc{camchoreo,
  title         = {Temporally Grounded Compositional Camera Motion Understanding via
                   Geometric Knowledge Distillation},
  year          = {2026},
  eprint        = {2608.10932},
  archivePrefix = {arXiv}
}

Citation

If you use this repository, please cite it. The upstream works credited above should be cited as well.

@misc{lin2026longformbench,
  author       = {Zeteng Lin and Jing Tang},
  title        = {Long-form Camera Motion Bench},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Zeteng/longform-camera-motion-bench}}
}

Downloads last month: 38

Total file size:

31.1 MB

Papers for Zeteng/longform-camera-motion-bench

Temporally Grounded Compositional Camera Motion Understanding via Geometric Knowledge Distillation

Paper • 2608.10932 • Published Aug 11

Natural Language Camera Movement Understanding

Paper • 2607.03043 • Published Jul 3