Research map / Image generation / Autoregressive and latent generation
Aurelius: Relation Aware Text-to-Audio Generation At Scale
Yuhang He, He Liang, Yash Jain, Andrew Markham, Vibhav Vineet
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 |
|---|---|---|---|
| finetune on the curated 100 h relation-aware data (ablation) | TangoFlux · Aurelius test set | mAMSR: 1.77 -> 5.58 | Improved |
| finetune on the curated 100 h relation-aware data (ablation) | Tango 2 · Aurelius test set | mAMSR: 1.29 -> 2.86 | Improved |
| finetune on the curated 100 h relation-aware data (ablation) | Tango · Aurelius test set | mAMSR: 0.79 -> 2.73 | Improved |
| finetune on the curated data instead of training from scratch (ablation) | TangoFlux · Aurelius test set | mAMSR: 2.58 -> 5.58 | Improved |
| finetune on the curated data instead of training from scratch (ablation) | Tango · Aurelius test set | mAMSR: 2.64 -> 2.73 | No clear effect |
| increase training data from 100 h to 300 h with finetuning (ablation) | TangoFlux · Aurelius test set | mAMSR | No clear effect |
| increase training data from 100 h to 300 h when training from scratch (ablation) | TangoFlux · Aurelius test set | mAMSR | Improved |
| use a Qwen2.5 32B agent to decompose the prompt into single events before TangoFlux | TangoFlux · Aurelius test set (zero-shot) | mAMSR: 1.77 -> 0.60 | Hurt |
| use a Qwen2 7B agent to decompose the prompt into single events before TangoFlux | TangoFlux · Aurelius test set (zero-shot) | mAMSR: 1.77 -> 0.04 | Hurt |
| move from single-event prompts to multi-event relation-aware prompts | TangoFlux · Aurelius test set | event accuracy: 75% -> 12% | Hurt |
Closest papers by meaning
- AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation (NeurIPS 2026 · similarity 0.87)
- AudioGenX: Explainability on Text-to-Audio Generative Models (AAAI 2025 · similarity 0.87)
- MATS: An Audio Language Model under Text-only Supervision (ICML 2025 · similarity 0.86)
- Unified Synthesis of Compositional Speech and Sound from Free-Form Text Prompts (NeurIPS 2026 · similarity 0.86)
- ETTA: Elucidating the Design Space of Text-to-Audio Models (ICML 2025 · similarity 0.86)
- TTA-Bench: A Comprehensive Benchmark for Evaluating Text-to-Audio Models (AAAI 2026 · similarity 0.86)
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.
