Google's Lyria 3.5 lets you fix one part of an AI song without redoing the whole thing
The new music model adds Selective Section Painting, so a bad chorus no longer means regenerating the entire track. Google still will not say what it trained on.
Google has released Lyria 3.5, the newest version of its music generation model, and the headline feature is not the sound quality. It is the ability to change your mind. A new function called Selective Section Painting lets you edit one specific part of a track, a verse, a chorus, a drum fill, while leaving everything else untouched.
Anyone who has used an AI music tool knows why that matters. Until now the workflow was closer to a slot machine than to editing: you typed a prompt, you got a whole song, and if the bridge was wrong you rolled again and hoped the rest survived. Section painting turns that into something more like working with layers in a photo editor.
Google says the model itself also improved. Melodies sound more natural, lyrics come out better, and vocals are more realistic with clearer pronunciation. You get finer control over tempo and track length, and you can fine tune the tempo and duration of individual elements such as vocals, drums and bass separately. Tracks can run from 30 seconds up to 3 minutes. Short melody sketches can be grown into full songs rather than thrown away. Lyria 3.5 is available now through Google Flow Music.
The interesting shift here is not audio fidelity, it is editability. Generative models are naturally all or nothing: the system produces one finished artifact from one prompt, and small corrections are expensive because everything is entangled. Giving users a way to regenerate just a slice, while keeping the rest coherent, is a hard technical problem and a sign the field is maturing from demo toys into actual tools. The same pattern showed up in image generation a couple of years ago, when inpainting turned generators into something designers would actually open twice.
One thing Google is still not saying is what the model learned from. When Lyria 3 launched, the company said it had trained on material that YouTube and Google had the right to use under their terms of service, partner agreements and applicable law, without naming specifics. Asked about the training data for Lyria 3.5, Google reportedly did not immediately respond. For a technology built on top of other people’s music, that silence is worth noting.
What this means for you: If you have never generated music before, this is a good moment to try, because the frustrating part, being stuck with a nearly good track, is the part that just got fixed. If you already make music with AI, section painting is the difference between a novelty and a draft you can iterate on. And if you write or perform music yourself, the training data question is the one to keep an eye on. Better tools built on unclear foundations remain unclear foundations.
Sources
Source: https://blog.google/innovation-and-ai/models-and-research/google-labs/lyria-3-5/
OpenAI open-sources its vulnerability-hunting tool, and hands it to anyone with a terminal
Codex Security CLI is Apache 2.0 licensed, scans repositories, verifies fixes and plugs into CI pipelines. It is the same system that helped patch 3,000 critical vulnerabilities.
