• Amazon launches Titan Image Generator, allowing AWS users to customize images using text descriptions.
  • Titan Image Generator incorporates built-in mitigations for toxicity and bias, addressing concerns in AI ethics.
  • Amazon adds a default "tamper-resistant" invisible watermark to generated images, enhancing accountability and responsible AI practices.

Amazon introduced Titan Image Generator during AWS re: Invent 2023, allowing AWS customers to generate and customize images using text descriptions. 

Part of the Titan family of generative AI models, the image generator enables tasks such as seamlessly changing backgrounds and creating lifestyle images while retaining the main subjects. Amazon emphasizes the protection of customers against copyright accusations, aligning with its AI indemnification policy.

AI Image Generation with built-in mitigations for toxicity and bias

Titan Image Generator, trained on diverse datasets across various domains, offers built-in mitigations for toxicity and bias. While the effectiveness of these mitigations remains to be tested, Amazon highlights its careful approach to training models and data selection. 

The company also implements a "tamper-resistant" invisible watermark by default to address concerns about AI-generated misinformation and abuse imagery.

Watermarking and Amazon's commitment to AI ethics

In line with a voluntary commitment to AI ethics, Amazon's Titan Image Generator automatically adds a "tamper-resistant" invisible watermark to generated images. 

The watermark addresses the spread of AI-generated misinformation and aligns with Amazon's commitment to responsible AI practices. Amazon has yet to clarify details about the watermarking technique and its detectability by external tools.

Edited by Shruti Thapa