Fast Text-to-Audio Generation with Adversarial Post-Training

TL;DR

This paper introduces ARC post-training, an adversarial acceleration method for diffusion/flow models, significantly reducing text-to-audio generation latency without distillation, enabling near real-time performance on high-end and edge devices.

Contribution

The paper presents ARC post-training, a novel adversarial acceleration technique for diffusion/flow models that improves inference speed for text-to-audio generation without relying on distillation.

Findings

Generates 12 seconds of stereo audio in 75ms on H100 GPU

Achieves 7 seconds of audio generation on a mobile device

First adversarial acceleration method for diffusion/flow models not based on distillation

Abstract

Text-to-audio systems, while increasingly performant, are slow at inference time, thus making their latency unpractical for many creative applications. We present Adversarial Relativistic-Contrastive (ARC) post-training, the first adversarial acceleration algorithm for diffusion/flow models not based on distillation. While past adversarial post-training methods have struggled to compare against their expensive distillation counterparts, ARC post-training is a simple procedure that (1) extends a recent relativistic adversarial formulation to diffusion/flow post-training and (2) combines it with a novel contrastive discriminator objective to encourage better prompt adherence. We pair ARC post-training with a number optimizations to Stable Audio Open and build a model capable of generating $\approx$12s of 44.1kHz stereo audio in $\approx$75ms on an H100, and $\approx$7s on a mobile edge-device, the fastest text-to-audio model to our knowledge.