Audio ControlNet for Fine-Grained Audio Generation and Editing

TL;DR

The paper addresses the challenge of achieving fine-grained, temporally precise, and signal-level control in text-to-audio generation. It proposes Audio ControlNet, with two designs (T2A-ControlNet and T2A-Adapter) that inject time-aligned control signals into a pretrained FluxAudio backbone, enabling controllable synthesis without backbone retraining. The results show that T2A-Adapter delivers state-of-the-art performance on AudioSet-Strong for event-level and segment-level F1 with only 38M additional parameters, and the framework extends to precise audio editing via T2A-Editor. By representing controls as unified temporal sequences and enabling multi-condition composition, the approach offers extensible, efficient control and editing for T2A models, with practical impact for sound design and multimedia production. The work provides release-ready code, models, and benchmarks to foster further research in controllable audio generation and editing.

Abstract

We study the fine-grained text-to-audio (T2A) generation task. While recent models can synthesize high-quality audio from text descriptions, they often lack precise control over attributes such as loudness, pitch, and sound events. Unlike prior approaches that retrain models for specific control types, we propose to train ControlNet models on top of pre-trained T2A backbones to achieve controllable generation over loudness, pitch, and event roll. We introduce two designs, T2A-ControlNet and T2A-Adapter, and show that the T2A-Adapter model offers a more efficient structure with strong control ability. With only 38M additional parameters, T2A-Adapter achieves state-of-the-art performance on the AudioSet-Strong in both event-level and segment-level F1 scores. We further extend this framework to audio editing, proposing T2A-Editor for removing and inserting audio events at time locations specified by instructions. Models, code, dataset pipelines, and benchmarks will be released to support future research on controllable audio generation and editing.

Figures (9)

• Figure 1: An illustration of fine-grained text-to-audio (T2A) generation tasks and fine-grained audio editing tasks, along with the proposed Audio ControlNet.
• Figure 2: The architecture of FluxAudio, which consists of a hybrid backbone that integrates MMDiT and DiT, and is trained with a flow matching objective.
• Figure 3: The framework of T2A-ControlNet and T2A-Adapter.
• Figure 4: Conditions and their feature extractors.
• Figure 5: Architecture of T2A-Editor. Reference audio and editing event information are injected via an external network to extend the text-to-audio model for the audio editing task.