junha1125/openvid-frame-sequences-1M

OpenVid Frame Sequences — 1M adjacent frame pairs

Short, single-shot frame sequences cut from OpenVid-1M, built to train and evaluate models on what changes between two frames half a second apart.

One sample = 10 consecutive frames, 0.5 s apart (a 4.5 s span) → 9 adjacent frame pairs.

[f00] --0.5s--> [f01] --0.5s--> [f02] ... [f09]
      ^ the thing you describe / predict

Sequences	116,596
Frames per sequence	10 (0.5 s apart, t = 0.0 … 4.5 s)
Adjacent frame pairs	1,049,364
Source clips	87,002 (at most 2 sequences per clip)
Images	1,165,960 JPEGs, 768 px wide, q=3, mean 44 KB
Download size	~53 GB (frames) + 145 MB (index)

Source videos are not included — see OpenVid-1M for those. Every sequence carries the clip name and timestamps, so the source frames are recoverable.

Why this exists

Video-caption datasets describe a whole clip. Training a model to reason about change needs pairs that are close enough in time to share a scene, but far enough apart that something actually happened. This dataset fixes that interval at 0.5 s and filters for clips where motion is real, so that a difference caption between f_i and f_i+1 is never "nothing changed".

Measured on a uniform random sample: mean absolute pixel change between adjacent frames is 16.6 / 255 (median 15.5), and only 1.0% of sequences are effectively static.

Selection criteria

Applied to the full OpenVid-1M metadata (1,019,957 clips) before downloading a single byte:

motion score      ∈ [5.27, 60]     # 5.27 = top 29.5% of the candidate pool
aesthetic score   ≥ 5.0
duration          ≥ 4.75 s (one window) / ≥ 9.75 s (two windows)
sequences/clip    ≤ 2               # so no single video dominates the data
caption           must not match  talk|talking|interview|podcast
clip name         must not start with  celebv_  or  pixabay_
camera motion     unconstrained

Capping at 2 sequences per clip matters: without it a handful of long videos contribute hundreds of near-identical sequences each.

Camera motion is deliberately not constrained — pans and zooms are themselves valid visual change. The resulting mix is 63.4% static / 36.6% moving (pan_right 7.8%, pan_left 7.0%, undetermined 6.5%, zoom_in 2.8%, tilt_up 2.2%, …).

Composition

camera	static 63.4% · moving 36.6%
source resolution	64.9% from 1080p originals (all resized to 768 px wide)
motion score	min 5.27 · p25 6.76 · median 9.74 · p75 20.95 · max 60.00
source clip length	median 9.8 s
windows per clip	1 window: 57,408 clips · 2 windows: 29,594 clips

Files

frames/frames-000.tar … frames-063.tar   64 shards, ~850 MB each, 53 GB total
sequences.jsonl                          116,596 lines — the index
meta/manifest.csv                        111,536 rows: the full filtered candidate list
meta/part_index.jsonl                    1,019,957 rows: clip -> (OpenVid zip part, member, bytes)
scripts/                                 the pipeline that produced this
docs/                                    build runbook and result notes (Korean)

Each tar extracts to frames/<2-hex shard>/<clip>__w<N>/f00.jpg … f09.jpg, which is exactly the path written in sequences.jsonl. Extract every shard into the same directory and the index paths resolve as-is.

The 2-hex sharding (md5(sequence_id)[:2]) keeps any one directory under ~500 entries, which matters on network filesystems.

Getting it

pip install huggingface_hub
hf download junha1125/openvid-frame-sequences-1M --repo-type dataset --local-dir openvid_seq
cd openvid_seq && for t in frames/*.tar; do tar -xf "$t"; done

A line of sequences.jsonl

{
  "id": "<clip>__w0",
  "video": "<clip>.mp4",
  "frames": ["frames/90/<clip>__w0/f00.jpg", "...", "frames/90/<clip>__w0/f09.jpg"],
  "timestamps": [0.0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5],
  "step_sec": 0.5,
  "n_frames": 10,
  "video_caption": "the original OpenVid caption for the whole clip",
  "meta": {"seconds": "...", "fps": "...", "motion": "...",
           "aesthetic": "...", "camera": "...", "hd": "1"}
}

Window w frame i is the source video at w * 5.0 + i * 0.5 seconds. video_caption is the OpenVid caption for the entire clip, not for this window — it is useful as context for a captioning prompt, not as a label.

import json
seqs = [json.loads(l) for l in open("sequences.jsonl")]
pairs = [(s["frames"][i], s["frames"][i+1], s) for s in seqs for i in range(9)]
len(pairs)   # 1,049,364

Integrity

Every one of the 1,165,960 frame paths in sequences.jsonl was stat-checked against disk: 0 missing, 0 orphan JPEGs, 0 duplicate ids, 0 clips over the 2-sequence cap, all sequences exactly 10 frames at exactly 0.5 s spacing, no sequence below the motion floor. The 64 tars hold 1,282,812 entries = 1,165,960 JPEGs + 116,596 sequence dirs + 256 shard dirs.

Known limitations

• Shot changes: ~3.2% of sequences. OpenVid's scene segmentation is not perfect, so a small fraction of sequences contain a cut. If a pair looks like two unrelated scenes, drop it. This is measurable — a single adjacent pair whose pixel delta is several times the sequence's own median is almost always a cut.
• Residual talking-head footage. The caption filter only matches talk|talking|interview|podcast, so clips where someone talks to the camera without those words in the caption (driving-and-talking, lecture footage) survive. Tighten TALKING in scripts/select_clips.py if that hurts your use case.
• video_caption describes the whole source clip, so it can mention things not visible in a given 4.5 s window.
• Frames are re-encoded JPEG at q=3 from already-compressed video; fine detail is lossy.

Reproducing / extending

scripts/ contains the full pipeline. meta/manifest.csv is fixed-order with a cum_gb column, so raising the byte budget downloads only the next slice:

python3 scripts/fetch_clips.py --budget-gb 800 --workers 32   # this release used 600
python3 scripts/extract_frames.py --workers 176               # existing windows are skipped
python3 scripts/verify_dataset.py --full

budget	clips	sequences	adjacent pairs
600 GB (this release)	84,362	111,417	1,002,753
800 GB	96,176	128,540	1,156,860
1,000 GB	104,216	140,898	1,268,082
1,390 GB (all)	111,536	153,391	1,380,519

This release has slightly more than the 600 GB row because 2,640 clips from beyond the budget were already on disk and were extracted too.

Note for anyone pulling clips out of the OpenVid zips: zipfile.ZipFile.read() costs 3–4 HTTP round trips per member and Hugging Face throttles on request count, not bytes. scripts/hfzip.py:read_member() fetches the local header and payload in one Range request and parses it directly — 2.5× faster in practice (19 → 58 MB/s).

License and attribution

CC-BY-4.0, inherited from OpenVid-1M. These are frames extracted from OpenVid-1M clips; all credit for the underlying video collection and its captions goes to its authors.

@article{nan2024openvid,
  title   = {OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation},
  author  = {Nan, Kepan and Xie, Rui and Zhou, Penghao and Fan, Tiehan and
             Yang, Zhenheng and Chen, Zhijie and Li, Xiang and Yang, Jian and Tai, Ying},
  journal = {arXiv preprint arXiv:2407.02371},
  year    = {2024}
}

Intended for research use. The underlying clips are sourced from public video platforms; the OpenVid authors' terms and takedown process apply to them.