Amazon Launches Ethical Data Usage and Invisible Watermarking Titan Image Generator, Joining the AI Text-to-Image Competition. Amazon's foundational model ensemble, Titan, has welcomed a new member: the Titan Image Generator. This week, at the re:Invent conference in Las Vegas, Vice President of Analytics and Machine Learning, Swami Sivasubramaniam, announced the Titan Image Generator, stating that the tool is now available in preview. He further added, "You can customize these images with your own data to create content that better reflects your industry or brand." Amazon has entered the market of text-to-image models, competing alongside Adobe Firefly. While it is still early to evaluate, Titan's competitors have already faced challenges. For instance, Google's Imagen has acknowledged encoding biases that sometimes result in racial discrimination or toxic outputs. Other tools like DALL.E and Stable Diffusion have also observed similar issues. It is said that the model was trained on a "diverse dataset," although Subramaniam declined to provide detailed information about the data source. During the announcement, Sivasubramanian stated, "We have been very careful about how we train our models and the data we use." Amazon AWS is the largest provider of computing power and data storage rentals. However, it lags behind OpenAI and Microsoft Bing Image Creator in launching its own text-to-image model-based product. Since the release of the basic version in April, Amazon's Titan series has added new models, including some that aim to generate text more affordably than OpenAI's latest version. While large tech companies remain cautious about releasing image models like Imagen and CM3leon to the public, Midjourney, RunwayML, Stability AI, and Stable Diffusion already have active user bases. Most of them are reluctant to release primarily due to security concerns and the risk of generating harmful, biased, and stereotypical images. Runway ML, Midjourney, Stability AI, and Stable Diffusion retain the right to prohibit users from creating harmful images, and their platforms do not process obscene instructions. Amazon has also done the same and established measures to prevent biases. It is said that the feature rejects unsafe topics and checks user input and output. In contrast, Amazon refuses to release the dataset on which the model was trained and only trusts that it has built-in mitigations against harmful content. At re:Invent, Swami Sivasubramaniam said, "The Titan Image Generator is trained on a diverse dataset so that you can create more accurate outputs." To combat intellectual property theft and differentiate AI-generated images from real ones, Amazon has added invisible watermarks to their outputs. In addition to creating new images, the Titan Image Generator allows users to isolate, extract, or integrate new components and edit images. The most useful applications include changing background settings or integrating items into lifestyle photos. Amazon is also betting on attracting other major model manufacturers to offer their software to AWS customers. Unlike existing platforms, their primary models focus on the B2B market. Swami Sivasubramaniam said in the company's release, "Generative AI is considered the most transformative technology of our time, and we are inspired by customers applying it to new opportunities and solving business challenges." Amazon also emphasizes the model's applicability to various fields such as e-commerce, advertising, and entertainment. For example, companies can use their proprietary image customization models to maintain a consistent visual style. Earlier this year, Amazon agreed to invest up to $4 billion in AI startup Anthropic. According to Sivasubramaniam, under this agreement, AWS customers can use Anthropic's Claude model, including one released last week. He also mentioned that Amazon offers an updated version of Meta Platforms' Llama model. "As customers integrate generative AI into their businesses, they turn to Amazon Bedrock for leading models, custom features, agent capabilities, and enterprise-grade security and privacy, providing a fully managed experience."