TADA! Tuning Audio Diffusion Models through Activation Steering
Abstract
Research reveals that specific attention layers in audio diffusion models control distinct musical concepts, enabling precise manipulation of audio features through activation steering.
Audio diffusion models can synthesize high-fidelity music from text, yet their internal mechanisms for representing high-level concepts remain poorly understood. In this work, we use activation patching to demonstrate that distinct semantic musical concepts, such as the presence of specific instruments, vocals, or genre characteristics, are controlled by a small, shared subset of attention layers in state-of-the-art audio diffusion architectures. Next, we demonstrate that applying Contrastive Activation Addition and Sparse Autoencoders in these layers enables more precise control over the generated audio, indicating a direct benefit of the specialization phenomenon. By steering activations of the identified layers, we can alter specific musical elements with high precision, such as modulating tempo or changing a track's mood.
Community
About
In this work, we investigate the internal mechanisms ๐ฌ of text-to-audio ๐ถ diffusion models, showing that various musical concepts (instruments ๐ธ, vocal ๐ค, tempo ๐, mood ๐ค) are controlled by a small, shared subset of attention layers. By combining these insights with activation steering methods, we provide a new tool for controllable audio generation.
Our contributions:
๐ท We construct a counterfactual prompt dataset for probing layer-level roles in audio diffusion models.
๐ We demonstrate that musical attributes are concentrated in just 2โ4 cross-attention layers across all three architectures.
๐ We leverage this localization to apply Contrastive Activation Addition (CAA) and Sparse Autoencoders (SAEs) for targeted activation steering.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders (2026)
- Learning a Generative Meta-Model of LLM Activations (2026)
- ConceptCaps: a Distilled Concept Dataset for Interpretability in Music Models (2026)
- Diffusion Timbre Transfer Via Mutual Information Guided Inpainting (2026)
- From Knots to Knobs: Towards Steerable Collaborative Filtering Using Sparse Autoencoders (2026)
- Interpretable and Perceptually-Aligned Music Similarity with Pretrained Embeddings (2026)
- Activation Steering for Masked Diffusion Language Models (2025)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper