TangoFlux

Why we included this project

TangoFlux turns a text description into up to 30 seconds of 44.1kHz audio, and its flow-matching transformer gets there in only a few sampling steps, which makes it fast compared with earlier diffusion-based text-to-audio models. That speed is the main draw for people making quick voiceover placeholders, game and video sound assets, or any pipeline that renders audio on demand from text. The repo ships checkpoints, the training setup, and a preference-optimization stage, along with the CRPO dataset and the scripts used to generate it, so teams can adapt the model to their own data or reproduce the whole pipeline rather than just loading a pretrained checkpoint. One caveat before you build on it: the distribution is research-only under the Stability AI Community License, so check whether that.