Introduce Gemma 3n preview, Veo 3, Imagen 4 & Veo 2 updates
New Google AI Models and Tools Advance Creative Media Creation and On-Device AI
Google has announced numerous new models and tools that emphasise advanced generative media production and on-device capabilities to make AI more accessible.
Gemma 3n
Gemma 3n preview, a powerful, mobile-first AI model, is important. Gemma 3n, the first open model with a new architecture, was developed with Qualcomm Technologies, MediaTek, and Samsung System LSI. This architecture is optimised for rapid, multimodal AI to enable truly private and intimate interactions on computers, tablets, and phones.
Gemma 3n's main features:
Compared to Gemma 3 4B, it reacts 1.5 times faster on mobile devices, has better quality, and uses less memory. Per-Layer Embeddings (PLE) reduces RAM usage and allows models with 5B and 8B parameter counts to work with a dynamic memory footprint of ONLY 2GB and 3GB.
MatFormer training allows a 4B active memory footprint model to natively incorporate a layered state-of-the-art 2B active memory footprint submodel. This lets you dynamically balance quality and performance. We add mix-and-match capability to the 4B model to dynamically construct submodels for certain use scenarios.
Privacy-First & Offline Ready: Local execution enables features that safeguard user privacy and work offline.
Gemma 3n understands text, sounds, and visuals well and has excellent video understanding. Audio features include high-quality Automatic Speech Recognition and Translation. Accepting interleaved data from multiple modalities lets it understand complex relationships.
It performs better overall, but best in Japanese, German, Korean, Spanish, and French.
Gemma 3n gives an early peek at Android and Chrome's architectural breakthroughs with the Gemini Nano generation, which will be powered by this same architecture and released later this year. Google AI Edge for on-device development and Google AI Studio for cloud-based research allow developers to preview Gemma 3n today.
Google is releasing new generative media tools and models in addition to on-device improvements. These aim to empower artists and producers by pushing media generation.
The updated and new creative tools are:
Veo 3: A revolutionary video-making model that records speech, traffic, and bird sounds. It works in text and image prompting, lip syncing, and physics. Enterprise users can use Veo 3 on Vertex AI, while US Ultra subscribers can use Gemini and Flow.
Veo 2 updates: Creator feedback led to object add/remove capability, camera controls for precise movements, outpainting to extend the frame, and state-of-the-art reference-powered video for creative control and consistency. Flow has reference-powered camera and video controllers.
Flow: A Veo-specific AI filmmaking tool for creatives. Flow lets users create cinematic scenes, clips, and narratives using Google DeepMind's most advanced models (Veo, Imagen, and Gemini). Users can describe photos in plain language and manage plot elements. Google AI Pro and Ultra US subscribers can now use Flow.
Imagen 4, the latest model, produces gorgeous images with great typography quickly and accurately. Imagen 4 supports aspect ratios up to 2k, has enhanced typeface and spelling, has excellent detail clarity, and works well in many styles. The Gemini app, Whisk, and Vertex AI offer it with Workspace apps like Slides, Vids, and Docs. A faster version is coming.
Lyria 2: This music creation paradigm now allows unlimited research and innovation. Businesses may employ Vertex AI and creators can use Lyria 2. AI Studio and API access MusicFX DJ's interactive music composition model, Lyria RealTime.
Google emphasises responsible AI development. SynthID watermarks will help identify Veo 3, Imagen 4, and Lyria 2 outputs as AI-generated. This watermark-based verification gateway, SynthID Detector, is unveiled today to help consumers recognise AI-generated content. As AI advances, Google is apprehensive of open models and strives to enhance protocols. It aims to unleash creativity and help creators realise their ideas.