Amazon has released a new version of its in-house image generation model, Titan Image Generator, for AWS customers tapping into the Bedrock generative AI platform. Dubbed Titan Image Generator v2, the new release includes several advanced features, said Channy Yun, principal developer advocate at AWS. With v2, it’s now possible to guide image creation with reference images, edit visuals, remove backgrounds, and generate variations, Yun said in a blog post.
“Titan Image Generator v2 can identify and segment multi-foreground objects intelligently,” Yun writes. “It supports generating color-conditioned images based on specific color palettes and uses image conditioning to refine your creation.”
It is conditioned on an image—providing a reference image, focusing on specific visual properties such as edges, contours, and structure. It can further be finetuned with reference images like a product or company logotypes to make sure there exists one unique look in an image generated. AWS has never revealed exactly what data the Titan Image Generator models are trained on, only that it consists of a mix of proprietary and licensed data. Like many other companies, AWS does not want to disclose details about its training datasets in order to protect competitive advantage and avoid potential intellectual property disputes.
Instead, AWS offers an indemnification policy to protect customers against situations whereby the Titan Image Generator v2 inadvertently reproduces a copyrighted training example, rather than transparency regarding the reproduction.
On the recent second-quarter earnings call, Amazon CEO Andy Jassy showed profound confidence in generative AI technologies like the Titan models, set against a backdrop of concerns related to enterprise adoption and quickly mounting expenses for training and deployment of such models. “Generative AI is going to grow very rapidly,” Jassy said, “and it is going to be built largely in the cloud from day one.”