TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization

TL;DR

TangoFlux delivers a fast, open-access text-to-audio generation framework based on rectified flow, achieving 30-second outputs in under 4 seconds on a single A40 GPU with 515M parameters. It introduces CRPO, an online, self-improving alignment loop that uses CLAP as a proxy reward to create and optimize audio preferences, outperforming static datasets. The combination of a VAE-based audio encoding, text and duration conditioning, and an efficient flow-based backbone yields state-of-the-art objective and human-evaluated performance, especially on multi-event prompts, while maintaining runtime efficiency. This work emphasizes the practicality of non-proprietary data for competitive TTA and provides open-source tooling to accelerate research in text-to-audio generation.

Abstract

We introduce TangoFlux, an efficient Text-to-Audio (TTA) generative model with 515M parameters, capable of generating up to 30 seconds of 44.1kHz audio in just 3.7 seconds on a single A40 GPU. A key challenge in aligning TTA models lies in the difficulty of creating preference pairs, as TTA lacks structured mechanisms like verifiable rewards or gold-standard answers available for Large Language Models (LLMs). To address this, we propose CLAP-Ranked Preference Optimization (CRPO), a novel framework that iteratively generates and optimizes preference data to enhance TTA alignment. We demonstrate that the audio preference dataset generated using CRPO outperforms existing alternatives. With this framework, TangoFlux achieves state-of-the-art performance across both objective and subjective benchmarks. We open source all code and models to support further research in TTA generation.

Figures (8)

• Figure 1: A depiction of the overall training pipeline of TangoFlux.
• Figure 2: The trajectory of CLAP score and KL divergence across the training iterations. This plot shows the stark difference between online and offline training. Offline training clearly peaks early, by the second iteration, indicated by the peaking CLAP score and increasing KL. In contrast, the CLAP score of online training continues to increase until iteration 4, while the KL divergence has a clear downward trend throughout.
• Figure 3: Winning and Losing losses of $\mathcal{L}_{\text{DPO-FM}}$ and $\mathcal{L}_{\text{CRPO}}$ at each iteration. Winning and Losing losses increase each iteration, as well as their margin.
• Figure 4: Comparison between $\mathcal{L}_{\text{DPO-FM}}$ and $\mathcal{L}_{\text{CRPO}}$ w.r.t. (a) CLAP$_\text{score}$, (b) FD$_\text{openl3}$, and (c) KL$_\text{passt}$ across iterations.
• Figure 5: Comparison of (a) CLAP and (b) FD Scores vs Inference Time for each model. Results are plotted for step counts of 10, 25, 50, 100, 150, and 200.