Video Coherence Rankings
Scored across three axes: temporal consistency, motion quality, and prompt fidelity. 800 generation tasks per model, evaluated frame-by-frame and holistically.
Current Rankings
Sorted by composite score. Video coherence methodology differs from image scoring: temporal frame consistency is the primary axis. See methodology for details.
| # | Model | Composite | Temporal | Motion | Fidelity | Tested |
|---|---|---|---|---|---|---|
| 1 | meridian-video-3 | 87.3 | 83.9 | 80.4 | Aug 10, 2026 | |
| 2 | drift-pro-2 | 82.4 | 79.1 | 76.8 | Aug 10, 2026 | |
| 3 | flowframe-v2 | 77.8 | 74.3 | 72.1 | Aug 10, 2026 | |
| 4 | luminate-motion-4 | 71.4 | 69.8 | 67.3 | Aug 3, 2026 | |
| 5 | scenecraft-2.5 | 67.1 | 64.2 | 62.8 | Aug 3, 2026 | |
| 6 | velox-gen-1 | 61.3 | 57.9 | 55.4 | Jul 27, 2026 | |
| 7 | nexus-clip-1.3 | 51.0 | 46.8 | 44.3 | Jul 27, 2026 |
How we score video coherence
Video evaluation requires a different framework than image scoring. Each model generates 800 short clips (4-10 seconds) from structured prompts describing motion, scene continuity, and object interaction. We deliberately include prompts that stress-test physical plausibility.
Composite score weights temporal consistency (45%), motion quality (35%), and prompt fidelity (20%). Temporal consistency measures frame-to-frame coherence using optical flow deviation. Motion quality scores naturalness using a panel-rated sample of 100 clips per model.
Full per-prompt clip scores and confidence intervals available via Pro tier. Methodology details on the Benchmarks page.
Access the full video evaluation dataset
Per-clip scores, temporal consistency traces, and motion quality breakdowns for all models in the current cycle.