VEFX-Bench v1.0: Video Editing Quality Benchmark

Frozen release: VEFX-Bench v1.0 — frozen on 2026-08-03. The immutable data and evaluator identifiers are listed under Reproducibility.

VEFX-Bench is a benchmark dataset for evaluating video editing model quality. It contains 300 original source videos with editing instructions across 9 task categories.

Dataset Structure

• 0000.mp4 through 0299.mp4: 300 original source videos (720p)
• prompts.json: editing instructions indexed by video number
• benchmark_meta.json: the v1.0 manifest with category labels and filename mapping

Usage

1. Download the frozen v1.0 revision listed below.
2. Apply your video editing model following each prompt.
3. Save edited videos as 0000.mp4 through 0299.mp4 (matching index).
4. Submit the zip to the VEFX-Leaderboard.

Public evaluator reference code is available in VEFX-Code.

Evaluation Dimensions

• IF (Instructional Following): how well the edit follows the instruction (1–4)
• RQ (Render Quality): visual quality and temporal consistency (1–4)
• EE (Edit Exclusivity): preservation of unrelated content (1–4)

Overall and Aggregation Order

Overall is the primary GeoAgg ranking metric; it is not the arithmetic average of IF, RQ, and EE. For item (j), first normalize each soft dimension score to ([0,1]):

[ i_j=\frac{IF_j-1}{3},\qquad r_j=\frac{RQ_j-1}{3},\qquad e_j=\frac{EE_j-1}{3}. ]

Then compute the item-level weighted geometric aggregate, with IF weighted twice:

[ g_j=1+3\left(i_j^2r_je_j\right)^{1/4}. ]

Finally, average the 300 item-level values:

[ \mathrm{Overall}=\frac{1}{300}\sum_{j=1}^{300}g_j. ]

Missing or failed items receive (IF_j=RQ_j=EE_j=1) before aggregation. This protocol is mean of per-item GeoAgg, not GeoAgg of the three dataset-level dimension means. The arithmetic mean of the three dimension averages may be shown separately as a diagnostic Mean, but it is not used for ranking.

Task Categories

These names and counts are taken directly from the frozen benchmark_meta.json manifest.

Category	Count
Attribute Editing	33
Camera Angle Editing	34
Camera Motion Editing	33
Creative Edit	33
Instance Editing	33
Instance Motion Editing	33
Quantity Editing	34
Style Editing	34
Visual Effect Editing	33
Total	300

Reproducibility

Artifact	Frozen identifier
Dataset revision	3bf997e7eb4fa0d0c2d56cce5ddfccc1dfbda235
benchmark_meta.json SHA-256	277d89f5cf23af4163fe6ee120654210f51d19c6dee2d3aa76a3248471d1f333
Reward model revision	a15a8dbe1b3eb07ee0919e8de059f170436ec9ff
model.safetensors SHA-256	c3c0d03f770f0a73631206821922213de75413bb517f7d7a2fd9ab1f2c38f59d

Download by immutable revision:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="xiangbog/VEFX-Bench",
    repo_type="dataset",
    revision="3bf997e7eb4fa0d0c2d56cce5ddfccc1dfbda235",
    local_dir="./vefx_bench_v1",
)

Citation

Coming soon.