VEFX-Bench - Instruction Following: leaderboard

VEFX-Bench (2026) compares instruction-guided video editing systems on curated source videos and editing prompts, scoring each edit with a reward model trained on human-annotated edits. This board is the coverage-adjusted instruction-following score on a 1 to 4 scale; higher is better.

Metric: Instruction-following score as a coverage-adjusted model-level estimate (inverse-propensity-weighted mixed-effects) on the 1-4 rubric scale over the 300 curated VEFX-Bench source-video and instruction pairs, each system's edited video scored by the authors' VEFX-Reward-32B evaluator from soft expected predictions; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Kling o13.04
2Kling 3.0 Omni3.03
3Runway Gen-4.52.82
4Seedance 2.02.81
5Luma ray 32.7
6Grok Imagine2.61
7UniVideo2.29
8Luma ray 22.04
9VACE (VEFX-Bench checkpoint unspecified)2.03
10Wan2.62.01

Interactive version: theaggregate.ai/benchmark?slug=vefx-bench-instruction-following · How It Works · Data refreshed daily, snapshot 2026-11-21.