Google DeepMind Launches Lyria 3.5 Music Model to Power Google Flow Music
Google DeepMind launched Lyria 3.5, its latest music-generation model, which is now powering Google Flow Music. Built on latent diffusion, the model generates full songs lasting 3 minutes, moving beyond the short audio clips produced by previous versions.
Google did not disclose the training data scale, specific data sources, or publish independent benchmarks. The updated model incorporates human and critic feedback during training and adds a SynthID watermark to all outputs, while the platform will now support dialogue-based editing, stem separation, music video generation through Veo, and distribution tools.
From the sources (3 posts)
@googledeepmindRT @googleflowmusic: Meet Lyria 3.5, the newest music model from @GoogleDeepMind, now powering Flow Music: 🎤 Vocals: More expressive, dyna…
RT @googleflowmusic: Meet Lyria 3.5, the newest music model from @GoogleDeepMind, now powering Flow Music: 🎤 Vocals: More expressive, dyna…
@maxforai刚刚， @GoogleDeepMind 发布了新一代音乐生成模型Lyria 3.5。 这个模型主要升级了四件事： - 音乐结构更自然复杂 - 歌词质量和指令遵循更好 - 人声更真实且更有情绪 - 同时可以更精确地控制歌曲时长与节奏。 它能够直接生成最长3分钟的完整歌曲，而不只是几十秒的音乐片段。 从模型卡里来看，Lyria 3.5依然采用latent diffusion，在时间音频latent上进行扩散生成。 训练数据是带有不同粒度文本标注的音频，后训练则包