Google releases Imagen 2, a video clip generator

Imagen 2 is an enhanced image-generating tool within Google's Vertex AI developer platform, capable of creating and editing images based on text prompts. It is a family of models that can render text, emblems, and logos in multiple languages and overlay these elements onto existing images such as business cards, apparel, and products. Imagen 2 also includes features like inpainting and outpainting for image modification and can generate short, four-second live videos from text prompts.

The new capabilities of Imagen 2 in Vertex AI include:

• Inpainting: This feature allows users to remove unwanted parts of an image.
• Outpainting: It enables the addition of new components to an image and the expansion of its borders to create a wider field of view.
• Text-to-live images: Imagen 2 can now create short live videos from text prompts, which are particularly aimed at marketers and creatives for uses such as GIF generation for ads.

SynthID is an approach developed by Google DeepMind that Imagen 2 employs to apply invisible, cryptographic watermarks to live images. These watermarks are designed to be resilient to edits, including compression, filters, and color tone adjustments. However, detecting these watermarks requires a Google-provided tool that is not available to third parties. This technology is intended to help address concerns around the potential creation of deepfakes.

Live images from Imagen 2 might not be competitive with other video generation tools for several reasons:

• Resolution: Currently, live images are in low resolution (360x640 pixels), which is less than what some competitors offer.
• Video Length: Imagen 2 generates four-second clips, whereas tools like Runway can produce 18-second clips.
• Customizability: Other tools, such as Stable Video Diffusion, offer greater customizability in terms of frame rate.
• Photorealism: Upcoming tools like OpenAI's Sora are expected to achieve higher levels of photorealism.

These factors suggest that Imagen 2's live images may not yet match the capabilities of existing video generation tools in the market.

Concerns around the training data for Imagen 2 include:

• Source of Data: Google has stated that the models are trained primarily on public web data, but the specific sources such as blogs, transcripts, and forums are not disclosed.
• IP-related Lawsuits: The details of the training data could potentially lead to intellectual property-related lawsuits.
• Creator Rights: There is uncertainty about whether creators whose work may have been included in the training process will have the option to opt out or receive compensation.

These issues highlight the complexity and potential legal implications associated with the data used to train generative AI models like Imagen 2.