Amazon Shows Up To The AI-Art Party With Its Own 'Titan Image Generator'
Amazon is going where every tech giant is headed — generative artificial intelligence in all its flavors. Following the announcement of its own AI chatbot targeted at businesses and the development of AI chips in the same vein as Microsoft, the company has announced another product that actually has a popular appeal. The latest from the e-commerce and cloud giant is Titan, a text-to-image generation model that aims to pull the same trick as OpenAI's Dall-E and other such tools based on models like Stable Diffusion.
Unsurprisingly, Titan is also aimed at enterprises with its own suite of tools, such as a hidden watermark system to avoid plagiarism, the ability to edit images using text prompts and refine the system using a local repository to ensure that the images follow a brand's signature design language. Amazon says Titan can generate "realistic images with inclusive attributes and limited distortions." These two continue to be a problem, even with the latest models such as Dall-E, where attributes such as eyes are visibly fuzzy and details bungled.
Amazon adds that its text-to-image generator is capable of processing intricate prompts involving multiple objects and generating a specified number of variations. "Large volumes and at low cost" are key to Amazon's Titan pitch, which the company says will find a place among content creators, ad businesses, online sellers, and the media landscape.
Impressive capabilities for business customers
Amazon says Titan offers a range of image editing capabilities, such as using a text prompt for automatic image modifications, thanks to its integrated segmentation model. This model is adept at both inpainting, which involves editing an image with a mask, and outpainting, which allows for background alterations or extensions. These are not too far off from what the likes of Adobe are trying to serve, but Amazon's implementation seems to be the simplest yet.
In the images above and below, take a look at how Titan accepts one-line commands to alter the background or replace the main subject in an image. It can even add or remove elements, such as the subject's body extension from the generated image.
Users have the flexibility to set the image size and choose how many image variations the model should produce. Moreover, the model can be tailored with custom in-house data to ensure that the generated images align with specific brand standards or to create images in a particular style. This can be achieved by refining the model with images from past marketing campaigns, for instance.
In comparison, OpenAI's Dall-E can't process local images for subsequent edits. The best you can do is create a custom GPT after purchasing a ChatGPT subscription and use it for prompts trained on a local training dataset. Titan text-to-image generator is limited to English and is now available in Amazon Bedrock.