LVO-Bench / README.md
tyxy's picture
Document LongSCOP acceptance at NeurIPS 2026
3efd097 verified
|
Raw History Blame Contribute Delete
7.96 kB
metadata
pretty_name: LVO-Bench
size_categories:
  - n<1K
tags:
  - arxiv:2611.12445
  - video
  - video-outpainting
  - long-video
  - reference-guided-generation
  - benchmark
configs:
  - config_name: default
    default: true
    drop_labels: true
    data_files:
      - split: test
        path:
          - '*.mp4'
          - metadata.jsonl

LVO-Bench: Long Video Outpainting Benchmark

LongSCOP has been accepted at NeurIPS 2026.

LVO-Bench is the evaluation dataset introduced in LongSCOP: Semantically Consistent Long Video Outpainting by Yu Tang, Zhiyuan You, Zhixiao Wang, Fanghua Yu, Qingyu Zhang, Xiang Yin, Chao Dong, and Jinjin Gu (2026). It is designed to assess long-horizon generation stability, temporal coherence, and semantic consistency in video outpainting, with particular attention to methods that retrieve reference frames from across a video's timeline.

Source paper: arXiv:2611.12445 · Venue: NeurIPS 2026 (accepted) · Dataset Viewer

Video outpainting expands the visible spatial extent of a video while preserving its observed content. Over successive temporal chunks, a model must maintain consistent scene layouts, object identities, and appearances, while avoiding flicker, accumulated errors, and generation collapse.

This repository provides 52 full-frame evaluation videos and their technical metadata in a single test split. The 62K-clip training collection described in the LongSCOP paper is a separate dataset.

Benchmark construction

Section 4.1 of the LongSCOP paper describes a benchmark curated from LongVideoBench and Video-MME, with human expert curation and a multi-stage filtering pipeline:

  1. Shot segmentation: split source videos into temporally coherent shots and retain segments long enough to support temporally distant reference retrieval.
  2. Temporal variation filtering: remove near-static footage with little useful visual change, as well as clips with excessively violent motion.
  3. Geometric validation: use optical flow and robust point tracking to select clips with stable horizontal camera motion, limited rotation, and small scale drift.
  4. Reference recoverability: check that distant frames can align with target frames and reveal useful content near the left or right borders.

The resulting benchmark emphasizes videos in which reference retrieval can supply meaningful information for horizontal outpainting. Its selection favors recoverable geometry and should be considered when interpreting results on other motion patterns or unseen regions.

Release statistics

The following statistics are measured from the MP4 files in this release:

Property Value
Evaluation videos 52
Split test
Resolution 848 × 480 pixels, all videos
Container / video codec MP4 / H.264
Total video size 149,875,894 bytes (149.88 MB)
Total duration 581.90 seconds (9.70 minutes)
Clip duration 5.47–29.70 seconds; mean 11.19 seconds
Total decoded frames 16,011
Frames per clip 137–890
Frame rates Approximately 23.976–30 FPS; retained per video

num_frames is counted from decoded frames, and durations and frame rates are read from the encoded video streams. Filenames retain their original identifiers; frame-count or FPS strings embedded in a filename may describe an earlier processing stage. Use the metadata fields for the released files.

Dataset Viewer and fields

Open the Dataset Viewer to play videos and inspect their metadata. The dataset uses Hugging Face's native VideoFolder format.

Field Description
video Full-frame evaluation video, exposed as a Hugging Face Video feature.
video_id Unique clip identifier, matching the MP4 filename without its extension.
width, height Frame dimensions in pixels.
fps Numeric frame rate.
fps_rational Frame rate as an exact rational string, such as 30000/1001.
num_frames Number of decoded video frames.
duration_seconds Encoded video-stream duration in seconds.
size_bytes MP4 file size in bytes.
codec Video codec name.

The root-level metadata.jsonl contains one record per video. Its file_name field links each record to its MP4 and is converted into the video column by VideoFolder. This release contains full-frame videos and technical metadata; cropped inputs, masks, captions, and reference-frame annotations are not included.

Load with Hugging Face Datasets

Install a recent version of datasets with video support. Decoding frames also requires a compatible TorchCodec/PyTorch installation and FFmpeg; see the video loading guide.

from datasets import Video, load_dataset

dataset = load_dataset("tyxy/LVO-Bench", split="test")

# Inspect metadata and file paths without decoding video frames.
dataset = dataset.cast_column("video", Video(decode=False))
example = dataset[0]
print(example["video_id"], example["num_frames"], example["duration_seconds"])
print(example["video"]["path"])

To decode and sample a frame with a compatible video backend:

decoded = dataset.cast_column("video", Video())
frame = decoded[0]["video"].get_frame_at(0)
print(frame.data.shape)  # (channels, height, width)

Download the original files

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="tyxy/LVO-Bench",
    repo_type="dataset",
    local_dir="LVO-Bench",
    allow_patterns=["*.mp4", "metadata.jsonl", "README.md"],
)

Evaluation

Use these full-frame videos as targets and derive observed inputs and exterior masks under an explicitly documented outpainting protocol. Preserve temporal alignment between predictions and targets, and report the crop geometry, mask ratio, reference-selection policy, and frame sampling alongside results.

The LongSCOP paper reports PSNR and SSIM (higher is better), LPIPS and FVD (lower is better). Section 4.2 specifies frame values in [0, 1] for PSNR, SSIM, and FVD, and [-1, 1] for LPIPS. FVD uses 16 uniformly sampled frames per video. For reproducibility, also specify whether frame-level metrics cover the complete frame or only the generated region.

Reference and source attribution

The benchmark motivation and construction description above follow LongSCOP: Semantically Consistent Long Video Outpainting (arXiv:2611.12445), especially Sections 4.1–4.2. The numerical release statistics are measured from the files hosted here.

If you use LVO-Bench, please cite the LongSCOP paper:

@misc{tang2026longscop,
  title         = {LongSCOP: Semantically Consistent Long Video Outpainting},
  author        = {Yu Tang and Zhiyuan You and Zhixiao Wang and Fanghua Yu and Qingyu Zhang and Xiang Yin and Chao Dong and Jinjin Gu},
  year          = {2026},
  eprint        = {2611.12445},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  note          = {Accepted at NeurIPS 2026},
  url           = {https://arxiv.org/abs/2611.12445}
}

Please also acknowledge the source benchmarks, LongVideoBench and Video-MME.

License information

A standalone license for this release has not been specified. Source videos remain subject to their original rights and applicable source-dataset terms; this dataset card does not grant additional rights to the underlying footage.