VSR-Extent

Benchmark of 3,175 video super-resolution (VSR) outputs from 5 VSR systems on 635 source clips, with crowd labels for artifact presence and two binary extent attributes: spatial (small vs. large) and temporal (persistent vs. momentary). 118k retained worker votes from 1,103 workers, ≥15 per clip.

Anonymous submission for NeurIPS 2026 Evaluations & Datasets Track.

Repository layout

Path	Contents
data/items.parquet	Per-clip metadata, all pools (3,385 rows)
data/assignments.parquet	Raw worker votes, 150,716 rows (~118k post-filter)
data/experts.parquet	Expert role confirmations for hidden-control and instruction items
data/metrics.parquet	Per-frame SR-quality metric values, 34,925 rows (3,175 × 11)
initial_videos/	823 mp4 stimuli, v6 / v6.1 pools
extension_videos/	2,378 mp4 stimuli, v7 / v7.1 pools
metadata.json	Croissant metadata with RAI fields

Annotation pools

The data comes from two subjective comparison studies:

• Initial (initial_videos/, v6 / v6.1): first round on REDS-hard.
• Extension (extension_videos/, v7 / v7.1): second round adding LDV3 and more methods.

items.parquet also lists pilot pools v3–v5. These have no mp4 files. The paper's 3,175-clip analysis subset is:

primary = items[
    (items.role == "main") &
    items.sr_method.notna() &
    items.input_hash.notna() &
    items.group_key.notna()
]

Question wording differs between pool versions. Pools v6+ use q2a ∈ {persistent, momentary} and q2b ∈ {small, large}. Earlier pools used different enumerations (e.g. v4 q2a ∈ {few_more_boxes, ...}).

Quick start

Load tables

import pandas as pd

items       = pd.read_parquet("data/items.parquet")
assignments = pd.read_parquet("data/assignments.parquet")
experts     = pd.read_parquet("data/experts.parquet")
metrics     = pd.read_parquet("data/metrics.parquet")

Load via 🤗 datasets

from datasets import load_dataset

items = load_dataset("vsr-extent-authors/vsr-extent", name="items")["train"]
votes = load_dataset("vsr-extent-authors/vsr-extent", name="assignments")["train"]

Download a stimulus

from huggingface_hub import hf_hub_download

row = items[items.role == "main"].iloc[0]
folder = "initial_videos" if row.pool_id.startswith("v6") else "extension_videos"

path = hf_hub_download(
    repo_id="vsr-extent-authors/vsr-extent",
    repo_type="dataset",
    filename=f"{folder}/{row.mp4}",
)

Parquet schemas

items.parquet (3,385 rows)

Column	Type	Description
pool_id	string	Pool version: v3, v4, v5, v6, v6.1, v7, v7.1
mp4	string	mp4 filename. v6* → initial_videos/; v7* → extension_videos/
dataset	string	Source corpus: REDS, LDV3; pilot tags (easy, ...) for v3–v5
sr_method	string	RealBasicVSR, RealViformer, dove_vsr_x4, star_vsr, upscale_a_video_v2
group_key	string	Groups clips sharing one source clip across methods
input_hash	string	Source-clip hash; joins to metrics.parquet via (dataset, sr_method, input_hash)
role	string	main, hidden_control, or instruction
current_category	string	Canonical five-way category for controls
temporal_start_frame, temporal_end_frame, expected_frames, expected_fps, prepared_frames	int / float	Temporal crop metadata

assignments.parquet (150,716 rows)

Column	Type	Description
pool_id	string	Pool
mp4	string	FK to items.mp4
q1	string	Presence response
q2a	string	Temporal response (v6+: persistent / momentary)
q2b	string	Spatial response (v6+: small / large)
gold_q1, gold_q2a, gold_q2b	string	Expected answer for hidden-control assignments
has_gold	bool	Whether the assignment contained hidden controls
assignment_hash, suite_hash, worker_hash	string	Salted SHA-256 pseudonyms
duration_min	float	Assignment duration

experts.parquet (294 rows)

Column	Type	Description
pool_id	string	Pool
mp4	string	FK to items.mp4
role	string	Expert-assigned role
current_category	string	Expert-assigned category
manual_role_confirmed	bool	Manually confirmed by an expert

metrics.parquet (34,925 rows = 3,175 × 11)

Column	Type	Description
dataset	categorical	REDS, LDV3
metric	categorical	bd_jup, desra, dists, dover, ldl, lpips, ssim, ssm_jup, tlp, tof, vmaf
method	categorical	One of 5 VSR methods
cand, run	categorical	SR candidate and run hashes
input_hash	string	Joins to items.input_hash
frame_mean	list<float>	Per-frame mean of the metric's spatial map
frame_top1pct, frame_top5pct, frame_top10pct	list<float>	Per-frame top-k% aggregates (equal to frame_mean for scalar-per-frame metrics)
frame_support_area_frac	list<float>	Per-frame fraction of evaluated pixels
frame_bbox_cx_rel, frame_bbox_cy_rel	list<float>	Per-frame bbox centre, relative
frame_bbox_valid	list<int>	Per-frame bbox validity
rest_json	string	Metric-specific auxiliary fields (JSON: clip_scalars, histogram, per-metric extras)

A clip is keyed by (dataset, sr_method, input_hash). This triple joins items.parquet to metrics.parquet for the primary subset.

Aggregated labels

Aggregated labels are not stored. Compute them from assignments.parquet using the reproduction scripts (anonymous code repository in the OpenReview submission). Raw votes let users re-aggregate under alternative consensus rules.

Intended uses

• SR-quality metric evaluation (pooled and within-stratum AUC).
• VSR method comparison on matched source clips.
• Diagnosing metric sensitivity to spatial extent vs. temporal persistence.

Limitations

• Stimuli are central 50% spatial crops, 2 s long. Border or longer-context artifacts are out of scope.
• Extent attributes are binary; large covers both extended single-region and multi-region defects.
• Under majority-vote aggregation, the artifact-positive mass concentrates in the persistent-large cell. Raw votes retain small and momentary evidence; users can re-aggregate under softer consensus rules.
• Single crowd platform, Russian-language interface.
• Pool versions v3–v5 use different question enumerations from v6+.

Responsible use

Surveillance, identity attribution, and re-identification uses are disallowed. SR of low-resolution video can improve identifiability of people in source clips; do not repurpose for person-centred inference. Full RAI statement in metadata.json.

Licensing

• Annotations and derived labels: CC BY 4.0.
• Source corpora (REDS, LDV3) and VSR outputs retain upstream licenses; see the paper's licensing appendix and the code repository for per-component terms.

Citation

@misc{vsr-extent-2026,
  title  = {{VSR-Extent}: Measuring the Presence and Extent of
            Video Super-Resolution Artifacts},
  author = {Anonymous Authors},
  year   = {2026},
  note   = {Under review at NeurIPS 2026 Evaluations \& Datasets Track}
}