Google Just Dropped LYRIA 3: New AI Feature No One Expected

Google has officially launched Lyria 3, its newest music generation model, rolling it out within the Gemini app and powering Dream Track in YouTube's creator toolkit. This marks a significant step forward from previous research-only demos, making music generation directly accessible to millions of users. Lyria 3 allows users to generate 30-second music tracks using natural language prompts, describing genre, mood, tempo, and even lyric language.

A key advancement in Lyria 3 is its ability to automatically generate lyrics, vocals, and instrumentation, eliminating the need for manual input required by Lyria 2. Beyond text, Lyria 3 accepts images or videos as input, generating tracks that match the visual content. This positions music as a first-class modality alongside text and vision within the Gemini ecosystem, indicating audio is no longer an afterthought.

Lyria 3 generates audio at a 48 kHz sample rate using 16-bit PCM stereo output, delivering production-quality audio rather than compressed demo formats. While currently capped at 30 seconds in the Gemini app, the music's quality and complexity are notably higher, featuring full arrangements with multiple instruments and vocals. The model generates music from scratch, handling continuous and multi-layered elements like melody, harmony, rhythm, and long-range coherence.

Every track generated by Lyria 3 includes an imperceptible watermark created by Google's SynthID technology, embedded directly into the audio waveform. This watermark is detectable by software even after compression, slowing down, or re-recording, addressing copyright concerns and ensuring attribution. Users can verify the presence of SynthID by uploading tracks back into the Gemini app, providing a robust technical solution for digital signature within sound.

Google DeepMind introduced Lyria Realtime, a system that operates on a chunk-based auto-regressive stream, generating audio in two-second chunks over a bidirectional websocket connection. This allows for live steering using weighted prompts, enabling users to change mood or instrumentation while music plays, with control changes processed in under 2 seconds. This offers a dynamic and interactive creative experience distinct from static generation.

Additionally, Google has built the Music AI Sandbox, aimed at musicians and creators seeking more hands-on control. This sandbox allows users to transform simple hums or basic piano lines into full orchestral arrangements, use MIDI chords to generate vocal choirs, and change instruments with text prompts while preserving the melody. This fosters a 'human-in-the-loop' AI approach, where the model acts as a collaborative jamming partner.

Pomelli, Google's AI marketing experiment from Google Labs, is expanding its capabilities with a new feature called Photoshoot. Designed for small and medium-sized businesses, Photoshoot addresses the high cost of professional product photography. It allows businesses to upload a product image, select visual themes and templates, and generate polished marketing images that appear professionally shot.

These AI-generated images integrate directly into Pomelli's existing campaign flow, leveraging the platform's 'business DNA profile' to maintain brand identity and visual style. Photoshoot produces ready-to-use assets for social media, advertising, and campaigns without requiring users to leave the platform. This feature is clearly aimed at e-commerce sellers and small retailers, building on Pomelli's existing generative capabilities like animate through VO3.1 integration.

Stitch, Google's AI design tool (rebranded from Galileo AI), continues to expand its capabilities, including Figma export across all agents. A new agent named Hatter has appeared in development builds, described as capable of creating high-quality designs. The labeling as an 'agent' suggests it will handle complex, multi-step design tasks over time, potentially applying deeper reasoning to UI and layout generation, similar to a design-focused counterpart to DeepMind's DeepThink.

Stitch also introduces app store asset generation, allowing mobile app designers to automatically create store-ready screenshots, descriptions, and app icons, significantly saving time for indie developers. Furthermore, native MCP (Model Context Protocol) integration is being built into Stitch's export menu. This enables direct connection to coding tools like Cursor, Claude, and Gemini CLI via an API key, streamlining the workflow for designers and developers by minimizing friction in pulling designs into coding environments.