Amazon Enhances AI Image Generator for Better Results

Amazon’s updated Titan Image Generator v2, which is available on AWS’s Bedrock platform, lets you change and condition images in more ways. However, there are still concerns about how clear the training data is.

Amazon has made an updated version of its image-making model, Titan Image Generator, available to AWS customers who use the Bedrock generative AI platform. In a blog post, Channy Yun, AWS’s principal developer advocate, highlights the new model’s numerous features, Titan Image Generator v2.

Yun says that users can “guide” their images by using reference images, changing existing images, getting rid of backgrounds, and making different versions of images.

Yun writes, “Titan Image Generator v2 can intelligently find and segment multiple foreground objects.” You can use a colour palette to make color-conditioned images with the Titan Image Generator v2. Furthermore, the image conditioning feature allows you to shape what you create.

Titan Image Generator v2 can condition images and can optionally take in a reference image to focus on certain visual elements in that image, such as edges, object outlines, and structural elements.

You can also use reference images, like a product or company logo, to fine-tune the model and make sure that the images it creates look consistent.

There is still a lot of confusion at AWS about what data it uses to train its Titan Image Generator models. The company previously stated that it utilized a combination of its proprietary data and licensed data.

A lot of vendors don’t want to give out this kind of information. They keep training data and information about it very secret because they think it gives them an edge over their competitors. Training data details could also lead to intellectual property lawsuits, so don’t give away too much.

AWS Titan Models

However, instead of being open, AWS has an indemnification policy that protects customers if a Titan model, such as Titan Image Generator v2, regurgitates i.e., spits out a mirror copy of a training example that might be protected by copyright.

Amazon CEO Andy Jassy said on the company’s most recent earnings call that he is still “very bullish” on generative AI technology like AWS’ Titan models, even though the enterprise is starting to doubt them and the costs of training, fine-tuning, and serving models are going up. “It’s going to get big fast in the generative AI space,” he said. “We will build most of it in the cloud from the start.”