Amazon has joined the AI image generation arena with the launch of its Titan text-to-image AI model. This new Titan Image Generator was announced during the AWS re:Invent conference. It is capable of producing “realistic, studio-quality images” and is designed with built-in guardrails against toxicity and bias. Unlike standalone apps or websites, Titan is a tool that developers can utilize to create their own image generators powered by the model. To make use of it, developers will need access to Amazon Bedrock.
During his keynote at the conference, Swami Sivasubramanian, AWS vice president of database, analytics, and machine learning, showcased the capabilities of the Titan Image Generator. He emphasized its ability to not only create an image from a natural language prompt but also to change backgrounds. The model targets an enterprise audience, distinguishing itself from the more consumer-oriented focus of well-known existing image generators like OpenAI’s DALL-E.
All images generated by the Titan Image Generator will automatically include invisible watermarks, as per Amazon’s announcement. Vasi Philomin, vice president for generative AI at AWS, mentioned to The Verge that this aligns with the voluntary commitments Amazon made with the White House in July.
Philomin explained, “We wanted a way to mark an image as created with AI, specifically made with Titan Image Generator, without affecting the visual, experiencing no latency, and being unable to be cropped or compressed out.” He also clarified that the watermark is not limited to the file’s metadata.
However, the detection of the invisible watermark comes with a catch. Amazon has developed an API for people to connect and check the image’s origin. This design choice reflects Titan’s nature as a model rather than an end product. Consequently, developers using Titan Image Generator will have the freedom to decide how to provide that information to users.
Sivasubramanian also announced the general availability of other Titan models: Titan Text Lite, a smaller model suitable for lighter text generation tasks such as copywriting, and Text Express, intended for more extensive uses like conversational chat apps.
Amazon will also provide copyright indemnity to customers using its Titan foundation models, including text-to-image. It will extend legal cover to individuals using any Amazon-created AI application, even if the app utilizes a different foundation model from its Bedrock AI model repository, such as Meta’s Llama 2 or Anthropic’s Claude 2. These applications encompass AWS HealthScribe, CodeWhisperer, Amazon Personalize, Amazon Lex, and Amazon Connect Contact Lens.