Projects

AIBA — Attention-based Instrument Band Alignment

• Conference: NeurIPS 2025 Workshop (AI for Music)
• What it does: Aligns attention maps from text-to-audio diffusion models with instrument frequency bands, enabling interpretable control.
• Links: Workshop · arXiv · PDF

---

Jamendo-QA — Large-Scale Music Question Answering Dataset

• Conference: Submitted to ICASSP 2026
• What it does: Builds a large-scale QA dataset from Jamendo music tracks, combining captions, tags, and questions for training multimodal LLMs.
• Links: arXiv · PDF · HuggingFace

---

MINO (Music-DINO)

• Status: Research in progress
• What it does: Adapts DINO self-distillation to music by using CQT spectrograms, harmonic-aware positional encoding, and dual-axis attention to capture pitch, harmony, and tempo.
• Links: (coming soon: code + paper)

---

Let Triggers Control — Frequency-aware Dropout

• Status: Under Review
• What it does: Introduces a frequency-aware dropout method for token control, enabling better handling of trigger tokens in generative models.
• Links: (preprint link coming soon)

---

Illustrious — Open Advanced Illustration Model

• Status: Technical Report (2024)
• What it does: Large-scale illustration generation model with open release for research and creative use.
• Links: arXiv · PDF · HuggingFace · CivitAI

---

Contrastive Adapter Training (CAT) — Personalized Image Generation

• Conference: CVPR 2024 Workshop (Generative Models for Computer Vision)
• What it does: Proposes a contrastive adapter training strategy to personalize diffusion models.
• Links: Workshop · arXiv · PDF