What is Google Lyria 3 Pro and how is it used to generate music for AI videos?

Lyria 3 Pro is Google DeepMind's music generation model, built for longer, structurally coherent compositions — full tracks up to roughly three minutes with a real musical arc rather than looping fragments. For AI video, you prompt it with mood, instrumentation, and an energy structure matched to your scenes, then sync the track to your cut.

Lyria 3 Pro sits in Google's generative stack alongside Veo for video: it accepts text (and image) prompts and returns structured music — intro, build, resolution — instead of short ambient loops, which is what makes it usable as an actual soundtrack rather than background texture. Google surfaces it through Vertex AI, the Gemini API, Google AI Studio, and Google Vids.

Prompting it for a video score. Write the prompt the way you'd brief a composer, mapped to your edit: genre and mood, instrumentation, tempo, and the energy arc tied to your scene structure — for example, "ambient piano intro for the establishing shots, percussion build into the chase, strings resolve on the final frame." Because the model understands song structure, a scene-aligned prompt produces a track whose peaks you can cut to, instead of one you fight in the timeline. Generate, download, drop it under your picture, and trim to your beats in the edit.

Using it inside an agentic video workflow. invideo is an agentic video creation tool with the current generation models available, and the invideo agent generates Lyria 3 Pro-style music scores as part of the same production run as your footage — so the score inherits the project's tone context instead of being prompted cold in a separate tool. In one documented production of a 45-second cinematic scene, the invideo agent delivered 26 video clips, 4 text cards, and 1 music score as a single downloadable folder — and generated 3 music score options proactively, in parallel with video generation, without being asked. Practically, that means you review score options the same way you review takes: pick one, reject all, or ask for a different register.

Licensing and watermarking. Lyria 3 Pro output is watermarked with SynthID — an inaudible identifier marking the audio as AI-generated. It doesn't change how the track sounds, but check Google's usage terms for your commercial or monetization context before publishing, since the watermark and license conditions travel with the audio.

Known limitations. User reports flag genre constraints — niche or hybrid styles can take multiple attempts to land — and generation quotas that cap how many tracks you can iterate through per period. Budget a few generations per scene rather than expecting a one-shot score. One boundary worth knowing: Lyria 3 Pro handles the music score; diegetic sound (footsteps, doors, ambience) is increasingly generated natively by the video model itself — in one short film, the creator manually added only 1 sound effect, with all other audio generated natively inside Seedance 2.0.

Watch some of these to see what works for you:

The magic moments are when Agent One does the step you didn't ask for.

— invideo's creative team