Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates.
MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction
Authors: Yunkee Chae, Kyogu Lee
NeurIPS 2025 | Venue PDF | LLM Run Details | Input Tokens: 36,047 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 2,985 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we outline our experimental protocol, including baseline models, datasets, and key implementation details. A comprehensive description of hyperparameters and training procedures is provided in Appendix D. All baseline results are re-evaluated using our test sets to ensure consistency with our experimental setup. |
| Researcher Affiliation | Academia | Yunkee Chae1 Kyogu Lee 1,2,3 Music and Audio Research Group (MARG) 1 Interdisciplinary Program in Artificial Intelligence (IPAI) 2 AIIS, 3 Department of Intelligence and Information Seoul National University EMAIL |
| Pseudocode | Yes | Algorithm 1 Inpainting using the Re Paint approach. Algorithm 2 Inpainting using adaptive timestep approach |
| Open Source Code | Yes | We released the full training code and all model checkpoints under an open-source license. |
| Open Datasets | Yes | We train and evaluate on three multi-track music datasets: Slakh2100 [37], MUSDB18 [38] (denoted Mu), and Moises DB [39] (denoted Mo). |
| Dataset Splits | Yes | Slakh2100... comprises 2100 songs divided into training (1500), validation (375), and test (225) splits... MUSDB18... use all 100 tracks from the official training split for training and the 50-track test split for evaluation... Moises DB... randomly sample 24 tracks (10%) as the test set and use the remaining tracks for training. |
| Hardware Specification | Yes | All models were trained on a single NVIDIA RTX 6000 GPU (48 GB of memory). |
| Software Dependencies | Yes | Text embeddings are obtained from CLAP [56] using the checkpoint music_audioset_epoch_15_esc_90.14.pt via the laion-clap library.5 Our implementation builds upon the official stable-audio-tools repository from Stability AI6 and the training framework from friendly-stable-audio-tools.7 |
| Experiment Setup | Yes | All of our models, except the one trained on the full dataset combination (SA, SB, Mu, Mo), are trained for 200K iterations with a batch size of 64, using 16 k Hz audio segments of 10.24 seconds. The full combination model is trained for 320K iterations with a batch size of 128. During sampling and inpainting, we apply classifier-free guidance (CFG) with a guidance scale of 2.0 and a per-track dropout probability of p = 0.1. All diffusion-based samples including those from baseline models are generated using 250 inference steps. |