TMD-Bench - Music Generation: leaderboard
Metric: Music score (0-100): the mean of four group scores, each normalized to 0-1 by the paper and reported here times 100: CLAP text-audio similarity; the mean of the four Audiobox-style aesthetic axes (production quality, production complexity, content enjoyment, content usefulness); the mean agreement of the paper's Music Captioner labels with six prompt attributes (instruments, rhythm, tempo, genre, ambiance, functional scene); and the mean of four Gemini 3 Pro MOS judgments, on the 100 prompts of the TMD-Bench test set (text-driven joint generation of music and a dance video); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Veo 3 | 54 |
| 2 | Kling 2.6 | 53 |
| 3 | Sora 2 | 52 |
| 4 | Seedance 1.5 pro | 52 |
| 5 | Ovi | 52 |
| 6 | JavisDiT | 46 |
| 7 | UniVerse-1 | 45 |
| 8 | LTX-2 | 45 |
| 9 | Wan 2.6 | 44 |
Interactive version: theaggregate.ai/benchmark?slug=tmd-bench-music-generation · How It Works · Data refreshed daily, snapshot 2026-10-07.