Research map / Image generation / Autoregressive and latent generation
PADS-TAL: Padding-Annealed Diffusion Sampling in Text-Aware Latent Space for Robust and Diverse Text-to-Music Generation
Taekoan Yoo, Wonkyung Jung, Kyunghun Kim, Kyeongbo Kong
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 |
|---|---|---|---|
| perturb only padding positions (PADS) instead of full conditioning (CADS), with audio latent (ablation) | SAO diffusion model · SongDescriber, matched CLAP-f | Vendi: 22.591 -> 24.367 | Improved |
| perturb only padding positions (PADS) instead of CADS, with audio latent (ablation) | SAO diffusion model · SongDescriber, matched CLAP-f | FDopenl3: 147.730 -> 139.838 | Improved |
| train diffusion in Text-Aware Latent space instead of audio latent, with CADS (ablation) | SAO + CADS · MelBench, matched CLAP-f | mRecall: 0.237 -> 0.391 | Improved |
| PADS in audio latent space without TAL (ablation) | SAO + PADS · MelBench, matched CLAP-f | mRecall: 0.237 -> 0.212 | Hurt |
| combine TAL and PADS versus audio latent with CADS (ablation) | SAO · SongDescriber / MelBench | Vendi: 22.591 -> 26.075 (15.4% higher diversity; 71.6% within-genre) | Improved |
| add PADS on top of TAL (versus TAL with CADS) (ablation) | TAL-DM · SongDescriber / MelBench | Vendi: 23.873 -> 26.075 | Improved |
| apply PADS instead of CADS to ACE-Step | ACE-Step · SongDescriber | CLAP-f: 0.1981 -> 0.2318 | Improved |
| apply PADS to MelodyFlow | MelodyFlow · SongDescriber | Vendi: 18.2190 -> 21.5843 | Improved |
| add CADS perturbation to SAO (full conditioning noise) | SAO · Human evaluation, alignment MOS | MOS alignment: 3.83 -> 2.87 | Hurt |
| use PADS instead of CADS on SAO | SAO · Human evaluation, within-model | MOS quality: 3.18 -> 3.89 | Improved |
| use PADS-TAL instead of SAO baseline | SAO · Human evaluation, across-model | MOS diversity: 3.41 -> 3.71 | Improved |
| use step annealing schedule instead of linear in PADS (ablation) | SAO + PADS · SongDescriber | CLAP-f: 0.335 -> 0.312 CLAP-f; Vendi 23.230 -> 24.843 | Mixed |
Closest papers by meaning
- ETTA: Elucidating the Design Space of Text-to-Audio Models (ICML 2025 · similarity 0.86)
- SoundCTM: Unifying Score-based and Consistency Models for Full-band Text-to-Sound Generation (ICLR 2025 · similarity 0.86)
- DITTO: Diffusion Inference-Time T-Optimization for Music Generation (ICML 2024 · similarity 0.85)
- Text Diffusion with Reinforced Conditioning (AAAI 2024 · similarity 0.85)
- MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction (NeurIPS 2025 · similarity 0.85)
- IMPACT: Iterative Mask-based Parallel Decoding for Text-to-Audio Generation with Diffusion Modeling (ICML 2025 · similarity 0.85)
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.
