Research map / Image generation / Autoregressive and latent generation
Simple and Controllable Music Generation
Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, Alexandre Défossez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the single-stage MusicGen text-to-music model | Text-to-music generation · MusicCaps Test Set | overall quality: 80.5 -> 84.8 | Improved |
| Add melody conditioning during training and supply chroma at test time | MusicGen 1.5B · In Domain Test Set | melody alignment: 64.44±0.83 -> 72.87±0.93 | Improved |
| Add melody conditioning (ablation) | MusicGen 1.5B · MusicCaps Test Set | FAD_vgg: 3.4 -> 5.0 | Hurt |
| Add melody conditioning (ablation) | MusicGen 1.5B · MusicCaps Test Set | overall quality: 80.74±1.17 -> 81.30±1.29 | No clear effect |
| Flatten all codebooks instead of using the delay pattern (ablation) | MusicGen 300M · In Domain Test Set | FAD_vgg: 0.96 -> 0.86 | Improved |
| Predict all codebooks at each time step in parallel instead of using the delay pattern (ablation) | MusicGen 300M · In Domain Test Set | FAD_vgg: 0.96 -> 2.58 | Hurt |
| Use partial flattening instead of the delay pattern (ablation) | MusicGen 300M · In Domain Test Set | FAD_vgg: 0.96 -> 1.32 | Hurt |
| Scale the model from 300M to 3.3B parameters (ablation) | MusicGen · In Domain Test Set | perplexity: 56.1 -> 46.1 | Improved |
| Scale from 1.5B to 3.3B parameters (ablation) | MusicGen · In Domain Test Set | overall quality: 81.9±1.4 -> 79.2±1.3 | Hurt |
| Replace T5 with Flan-T5 as text encoder (ablation) | MusicGen 300M · MusicCaps Test Set | overall quality: 85.04±1.23 -> 85.54±1.01 | No clear effect |
| Replace T5 with CLAP as text encoder (ablation) | MusicGen 300M · MusicCaps Test Set | overall quality: 85.04±1.23 -> 82.13±1.29 | Hurt |
| Add metadata to the text description with Condition Merging (ablation) | MusicGen 300M · MusicCaps Test Set | FAD_vgg: 3.68 -> 3.28 | Improved |
Closest papers by meaning
- SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation (ICML 2025 · similarity 0.91)
- Whole-Song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models (ICLR 2024 · similarity 0.89)
- MuseControlLite: Multifunctional Music Generation with Lightweight Conditioners (ICML 2025 · similarity 0.89)
- MuPT: A Generative Symbolic Music Pretrained Transformer (ICLR 2025 · similarity 0.89)
- MusicFlow: Cascaded Flow Matching for Text Guided Music Generation (ICML 2024 · similarity 0.89)
- MusicRL: Aligning Music Generation to Human Preferences (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
