SyncTrack: Rhythmic Stability and Synchronization in Multi-Track Music Generation

Hongrui Wang; Fan Zhang; Zhiyuan Yu; Ziya Zhou; Xi Chen; Can Yang; Yang Wang

cs.SDMar 1, 2026

SyncTrack: Rhythmic Stability and Synchronization in Multi-Track Music Generation

Authors: Hongrui Wang, Fan Zhang, Zhiyuan Yu, Ziya Zhou, Xi Chen, Can Yang, Yang Wang

Organizations: Department of Mathematics, The Hong Kong University of Science and Technology · State Key Lab of CAD&CG, Zhejiang University · Academy of Interdisciplinary Studies, The Hong Kong University of Science and Technology · State Key Laboratory of Nervous System Disorders, The Hong Kong University of Science and Technology ·

Abstract

Multi-track music generation has garnered significant research interest due to its precise mixing and remixing capabilities. However, existing models often overlook essential attributes such as rhythmic stability and synchronization, leading to a focus on differences between tracks rather than their inherent properties. In this paper, we introduce SyncTrack, a synchronous multi-track waveform music generation model designed to capture the unique characteristics of multi-track music. SyncTrack features a novel architecture that includes track-shared modules to establish a common rhythm across all tracks and track-specific modules to accommodate diverse timbres and pitch ranges. Each track-shared module employs two cross-track attention mechanisms to synchronize rhythmic information, while each track-specific module utilizes learnable instrument priors to better represent timbre and other unique features. Additionally, we enhance the evaluation of multi-track music quality by introducing rhythmic consistency through three novel metrics: Inner-track Rhythmic Stability (IRS), Cross-track Beat Synchronization (CBS), and Cross-track Beat Dispersion (CBD). Experiments demonstrate that SyncTrack significantly improves the multi-track music quality by enhancing rhythmic consistency.

Figures & tables

Figure 1: (a) Previous methods ( Mariani et al., 2023 ; Karchkhadze et al., 2025 ) leverage a unified model to learn the joint distribution of multi-track audio stems. (b) While our proposed SyncTrack incorporates both track-shared modules and track-specific modules for common and specific information between tracks.

Figure 2: a. Overall pipeline for SyncTrack. Training pipeline: We train a four-track latent diffusion model. Each track is perturbed based on $l$ -th signal-to-noise ratio. The model is optimized to predict the added noise $\epsilon\in\mathcal{N}(0,I)$ . More details are in Section 3.1. Inference pipeline: At test time, four-track latents are generated and then decoded into audio data. b. SyncTrack consists of input, mid, and output blocks, which contains track-specific modules and track-shared modules.

Figure 3: Illustration of the (a) track-shared module and (b) track-specific module. In (a), we leverage inner-track attention to capture the inner-track rhythmic stability and devise (c) two cross-track attention submodules to capture cross-track rhythmic stability and synchronization. In (b), we construct a learnable instrument prior to capture timbre and other track-specific features.

Metrics	MSDM	STEMGEN	JEN-1 Composer	MSG-LD	SyncTrack
FAD	6.55	4.3	4.04	1.31	1.26

Table 1: FAD

\downarrow

scores of mixture music.

Method	Bass	Drum	Guitar	Piano
SyncTrack	0.710	0.710	1.450	1.110
MSG-LD	1.050	0.980	1.830	2.040
MSDM	6.304	6.721	4.259	5.563

Table 2: Track-wise FAD

\downarrow

scores.

	Mixture				Drum		Guitar
Method	1	2	3	4	1	2	1	2
Ground Truth	4.2 $\pm$ 0.9	4.5 $\pm$ 0.6	4.7 $\pm$ 0.5	4.6 $\pm$ 0.6	3.0 $\pm$ 0.2	2.6 $\pm$ 0.7	2.9 $\pm$ 0.3	3.0 $\pm$ 0.2
SyncTrack	3.3 $\pm$ 1.0	3.5 $\pm$ 0.8	3.0 $\pm$ 0.9	3.9 $\pm$ 0.9	1.9 $\pm$ 0.3	2.1 $\pm$ 0.5	1.9 $\pm$ 0.4	1.8 $\pm$ 0.5
MSG-LD	1.5 $\pm$ 0.6	1.3 $\pm$ 0.5	1.8 $\pm$ 0.9	1.7 $\pm$ 0.8	1.2 $\pm$ 0.5	1.3 $\pm$ 0.6	1.2 $\pm$ 0.5	1.2 $\pm$ 0.4

Table 3: Subjective evaluation results.

Method	Bass	Drum	Guitar	Piano
Ground Truth	0.015	0.005	0.016	0.015
SyncTrack	0.021	0.011	0.024	0.023
MSG-LD	0.041	0.040	0.039	0.039
MSDM	0.050	0.036	0.034	0.046

Table 4: Track-wise IRS

\downarrow

scores.

Metrics	Ground Truth	SyncTrack	MSG-LD	MSDM
CBS ↑	0.5740	0.5206	0.3861	0.4694
CBD (mean)↓	0.2412	0.2681	0.3714	0.3127
CBD (std)↓	0.1578	0.2131	0.2642	0.2217
CBD (median)↓	0.2066	0.2258	0.3545	0.2811

Table 5: Cross-track synchronization metrics scores.

Model	Bass	Drum	Guitar	Piano	Mixture	Promotion
Backbone	5.234	3.081	6.012	6.170	2.570	50.97%
Backbone w/ $\textcircled{a}$	0.816	0.809	2.634	1.695	1.742	27.67%
Backbone w/ $\textcircled{a}$ + $\textcircled{b}$	0.632	0.758	2.367	1.359	1.627	22.56%
Backbone w/ $\textcircled{a}$ + $\textcircled{c}$	0.892	0.889	2.680	1.547	1.429	11.83%
SyncTrack-alternate	0.900	0.897	2.663	1.757	1.586	20.55%
SyncTrack-reorder	0.957	0.943	2.887	1.877	1.681	25.04%

Table 6: Ablation study of track-specific module and two types of cross-track attention.

Figure 4: Comparison of subjective ratings and objective metric scores.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Comparison of IRS across hyperparameter settings,

Figure A2: Comparison of CBS across hyperparameter settings.

Setting	Metrics	Ground Truth	SyncTrack	MSG-LD	MSDM
fps@100 tl@50	CBD (mean) $\downarrow$	0.2206	0.2589	0.3644	0.2923
	CBD (std) $\downarrow$	0.1693	0.2241	0.2822	0.2370
	CBD (median) $\downarrow$	0.1769	0.2058	0.3350	0.2424
fps@100 tl@100	CBD (mean) $\downarrow$	0.2143	0.2522	0.3829	0.3205
	CBD (std) $\downarrow$	0.1556	0.2101	0.2708	0.2329
	CBD (median) $\downarrow$	0.1762	0.2060	0.3679	0.2870

Appendix

Table A1: Comparison of CBD across hyperparameter settings.

Figure A3: Track-wise IRS scores.

Panel of Figure 15 — Figure A4: Cross-track synchronization metrics scores.

Figure A5: Case study of IRS for ground truth and generated samples.

Figure A6: Visualization of cross-track synchronization in ground truth and generated samples.

	Sample	CBD(mean)	CBD(std)	CBD(median)	CBS
Slakh2100	①	0.0702	0.0481	0.0709	0.5286
Slakh2100	②	0.0628	0.0677	0.0553	0.7857
Generated from baselines	③	0.2619	0.2512	0.1522	0.3629
Generated from baselines	④	0.3975	0.2387	0.3815	0.3041

Appendix

Table A2: Comparison of CBS and CBD for ground truth and generated samples.

Layer (type)	depth-idx	Input Shape	Output Shape	Param #
SyncTrack	-	[1, 4, 8, 64, 64]	-	128,122,516
- DiffusionWrapper	1-1	[1, 4, 8, 64, 64]	[1, 4, 8, 64, 64]	-
- UNetModel	2-1	[4, 16, 64, 64]	[4, 8, 64, 64]	16
- Sequential (time_embed)	3-1	[4, 128]	[4, 512]	-
- Linear	4-1	[4, 128]	[4, 512]	66,048
- SiLU	4-2	[4, 512]	[4, 512]	-

Appendix

Table A3: Network architecture of SyncTrack

Panel of Figure 20 — Figure A7: Distribution of three metrics on MSG-LD, SyncTrack, MUSDB18 and Slakh2100.

	CBS $\uparrow$	CBD(mean) $\downarrow$	CBD(std) $\downarrow$	CBD(median) $\downarrow$	IRS(B) $\downarrow$	IRS(D) $\downarrow$	IRS(G) $\downarrow$	IRS(P) $\downarrow$
MSG-LD	0.386	0.371	0.264	0.355	0.041	0.040	0.039	0.039
MSDM	0.469	0.313	0.222	0.281	0.050	0.036	0.034	0.046
SyncTrack	0.521	0.268	0.213	0.226	0.021	0.011	0.024	0.023
MUSDB18	0.512	0.264	0.197	0.223	0.017	0.007	0.029	0.015
Slakh2100	0.574	0.241	0.158	0.207	0.015	0.005	0.016	0.015

Appendix

Table A4: Scores on three metrics for the MUSDB18 and Slakh2100

	GT	B	BP	BGP	BDGP
CBS $\uparrow$	0.5740	0.5620	0.5576	0.5259	0.5206
CBD(std) $\downarrow$	0.1578	0.1902	0.1940	0.2118	0.2131
CBD(median) $\downarrow$	0.2066	0.2192	0.2188	0.2304	0.2258
CBD(mean) $\downarrow$	0.2412	0.2591	0.2579	0.2718	0.2681

Appendix

Table A5: CBS and CBD metrics across different generation scenarios.

Panel of Figure 23 — Figure A8: Distribution of three metrics on different segment length.

Segment length	CBS $\uparrow$	CBD $\downarrow$			IRS $\downarrow$
		mean	std	median	Bass	Drum	Guitar	Piano
10s	0.5740	0.2412	0.1578	0.2066	0.015	0.005	0.016	0.015
30s	0.5554	0.2344	0.1742	0.1922	0.0203	0.007	0.0298	0.0254
60s	0.5349	0.2471	0.1960	0.2018	0.0258	0.0121	0.0425	0.0366
full-song	0.4844	0.2660	0.2466	0.1995	0.0489	0.0273	0.0671	0.0710

Appendix

Table A6: Slakh2100 scores on different audio segment length

Music Generation

Music generation studies computational methods for composing, synthesizing, and editing music from symbolic, audio, textual, and multimodal inputs.

See all 107 papers

Qwen-Music Technical Report
Jul 13, 2026Jin Xu, Kangdi Wang, Ruibin Yuan +24Text-to-Music Generation Controllable Music Generation
Qwen Team
ViTex: Visual Texture Control for Multi-Track Symbolic Music Generation via Discrete Diffusion Models
Mar 2, 2026Xiaoyu Yi, Qi He, Gus Xia +1Music Generation Controllable Music Generation
School of EECS, Peking University · Music X Lab, MBZUAI · Courant Institute of Mathematical Sciences, New York University
SymphonyGen: 3D Hierarchical Orchestral Generation with Controllable Harmony Skeleton
Apr 28, 2026Xuzheng He, Nan Nan, Zhilin Wang +7Controllable Music Generation Symbolic Music Generation
Department of AI Music and Music Information Technology, Central Conservatory of Music · Frontier Institute of Science and Technology, and Interdisciplinary Research Center of Frontier Science and Technology, Xi’an Jiaotong University · University of Science and Technology of China +1

Jul 13, 2026J/K move · Enter open · S save

Qwen-Music Technical Report

Jin Xu, Kangdi Wang, Ruibin Yuan, Shun Lei+23

Qwen Team

Text-to-Music Generation Controllable Music Generation Music Generation

Open paper AskPDF arXiv

We introduce Qwen-Music, a music generation model that produces high-fidelity songs with complete vocals. It supports text-to-music generation from descriptions, lyrics, and musical attributes, and cover song generation with different styles and vocal characteristics. Qwen-Music comprises three components: Qwen-Music-Tokenizer, Qwen-Music-LLM, and Qwen-Music-Render. The tokenizer compresses audio into a 25 Hz single-codebook stream of Music Semantic Tokens that preserve semantic and melodic information. The LLM performs autoregressive modeling with a melody-token-based chain-of-thought (Melody-CoT) mechanism that plans melodies before full-song generation, improving musicality, structural coherence, and reference-melody preservation. The renderer enriches discrete semantic tokens with acoustic details to produce high-fidelity stereo waveforms. We train the LLM using a quality-aware pre-training curriculum followed by progressive post-training with supervised initialization, offline DPO, and online GSPO to improve musicality and instruction following. On 600 evaluation inputs, Qwen-Music achieves state-of-the-art results in 13 of 16 objective musicality and audio-quality metrics. Professional evaluators also prefer Qwen-Music over leading proprietary systems. For cover generation, Qwen-Music preserves reference melodies more accurately than Suno V5.5, Suno V5, and MiniMax Cover on the AI-generated reference set, and outperforms MiniMax Cover on most metrics on the real-world popular-song reference set.

Abstract

Music Generation

Explore similar work

Qwen-Music Technical Report

ViTex: Visual Texture Control for Multi-Track Symbolic Music Generation via Discrete Diffusion Models

SymphonyGen: 3D Hierarchical Orchestral Generation with Controllable Harmony Skeleton