September 14th.Stability AI The enterprise level has been officially releasedAudio Generation Model Stable Audio 2.5, increased relative to the previous generation, mainly in terms of audio detail, production speed, with the name "Only 2 seconds to create 3 minutesAudio Track”.
According to a presentation, the core improvements of Stable Audio 2.5 focus on the ability to produce music, which is said to produce results that are more in tune with the actual chorus logicA full multi-band structure with foreplay, development and endI don't know. At the same time, the new model is more accurate in its understanding of the hints, particularly in terms of emotional description and soundness of the music-style vocabulary, and more responsive to expectations。
In addition, the new version of the model has significantly improved the speed of audio generation, which, according to Stability AI, is largely the result of a post-training approach proposed by the R & D team ARC (Note: Adversarial Relativistic-Contrastive), a technique that accelerates the production of the proliferation model by combining a relativist training and contrastorA SIGNIFICANT REDUCTION IN THE GPU REASONING TIME-CONSUMING WHILE ENSURING THE QUALITY OF THE TRACK CAN RESULT IN THE GENERATION OF AUDIO CONTENT OF UP TO 3 MINUTES IN 2 SECONDS.
In addition to this, Stable Audio 2.5 has added an audio patch that allows users to import their own audio files and assigns a "extend position"The model allows a "extend" to the audio, depending on the content of the sound and the overall curve, which is particularly suitable for a scenario such as a clip.
At present, Stable Audio 2.5 has been tested directly through the StableAudio network, while supporting localization. Officially, however, the audio files uploaded by users should not contain copyrighted content, and the StableAudio website will be tested using its own content identification system to ensure that copyrights are not violated。