FdAudio: MeanFlow-Anchored Fréchet-Distance Post-Training for One-Step Text-to-Audio Generation
Abstract
While recent few-step sampling text-to-audio generation models like MeanAudio substantially accelerate generation by modeling average velocities, their strict one-step generation quality still lags significantly behind multi-step counterparts. We propose FdAudio to bridge this gap. Unlike MeanAudio, which relies solely on regression against target velocity fields, our post-training approach optimizes the final one-step distribution directly across pre-trained embedding spaces via a multi-representation Fréchet-distance (FD) loss. Crucially, to prevent the multi-step degradation that naive post-training with FD-loss causes, we introduce a MeanFlow consistency objective as a structural anchor. Results demonstrate that FdAudio establishes state-of-the-art one-step T2A generation quality among few-step systems, yielding an 11.4% reduction in FD score and a 28.8% improvement in FAD score relative to the baseline MeanAudio framework. Notably, we solve FD post-training's naive multi-step degradation issue by proposing the MeanFlow anchor, enabling a 25-step sampling path to maintain high-fidelity audio synthesis that matches or surpasses strong multi-step models at a fraction of their computational latency.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
FdAudio presents a practical and effective post-training strategy for one-step text-to-audio generation, successfully bridging the quality gap with multi-step models while preserving sampling flexibility through MeanFlow anchoring. The work is technically sound, well-evaluated, and addresses a relevant problem in the field of generative audio.
Methodology
The paper addresses a critical issue in flow-matching based text-to-audio generation: the trade-off between one-step generation quality and multi-step sampling flexibility. The authors propose FdAudio, a post-training framework that combines Fréchet Distance (FD) loss with a MeanFlow consistency anchor. The core technical contribution is the identification that naive FD post-training collapses the velocity field, destroying multi-step sampling capability. By introducing a MeanFlow regularization term, they preserve the trajectory while optimizing the one-step output distribution. The methodology is sound and logically derived, leveraging existing concepts (FD loss, MeanFlow) in a novel combination. However, the novelty is incremental; it is an effective engineering solution rather than a fundamental theoretical breakthrough. The use of a queue-based estimator for FD loss is also a known technique in GAN/diffusion post-training.
Results & evaluation
The experimental evaluation is comprehensive and convincing. The authors demonstrate state-of-the-art one-step performance on AudioCaps, significantly outperforming MeanAudio and other few-step baselines. Crucially, they show that their method retains high-quality multi-step sampling, a feat that naive FD post-training fails to achieve. The ablation studies are thorough, analyzing the impact of the MeanFlow anchor weight, the choice of encoders for the multi-representation FD loss, and the comparison against various baselines. The inclusion of subjective human evaluation (MOS) alongside objective metrics (FAD, CLAP, IS) strengthens the claims. The latency measurements further validate the practical utility of the approach.
Reproducibility
The paper provides sufficient detail for reproduction. It specifies the base model (MeanAudio-S-Full), the training data (AudioCaps + WavCaps subset), the hyperparameters (learning rate, batch size, queue size), and the specific encoders used. The code and demo links are provided, which greatly enhances reproducibility. The training budget is mentioned as limited (1 epoch on 80k samples), which might affect generalizability, but the setup is clear.
Limitations
The primary limitation is the reliance on a limited training budget (1 epoch on a subset of data). While the results are strong, it is unclear if the method scales linearly with more data or longer training. Additionally, the method is specific to flow-matching/consistency models and may not generalize to other generative paradigms without adaptation. The improvement in multi-step quality is present but not revolutionary; the main gain is in one-step quality while maintaining multi-step capability.
Broader impact
This work has significant implications for real-time audio generation applications, such as interactive games, virtual assistants, and content creation tools, where low latency is critical. By enabling high-quality one-step generation without sacrificing the flexibility of multi-step sampling, it provides a more robust solution for diverse deployment scenarios. It also contributes to the broader understanding of post-training techniques for generative models, highlighting the importance of trajectory preservation. FdAudio presents a practical and effective post-training strategy for one-step text-to-audio generation, successfully bridging the quality gap with multi-step models while preserving sampling flexibility through MeanFlow anchoring. The work is technically sound, well-evaluated, and addresses a relevant problem in the field of generative audio.
Author of this paper? Send a correction.