Amazon makes Titan text models available and previews Titan Image Generator

Background

Amazon had announced its Titan models in April 2023, alongside the launch of Amazon Bedrock. Those first models were in limited preview. By late 2023 the company was extending the Titan family in several directions.^[1][2]

What happened

On 29 November 2023, during AWS re:Invent, AWS made Amazon Titan Text Express and Titan Text Lite generally available in Amazon Bedrock.^[1] Both are text models. Titan Text Express supports up to 8,192 tokens of context. Titan Text Lite supports 4,096 tokens and targets low-cost English tasks.^[1]

On the same day AWS released Titan Image Generator in public preview. Every image it generates carries an invisible watermark.^[1] The launch also brought in Titan Multimodal Embeddings, which is meant for text-and-image search and recommendation.^[1]

Key details

The Titan text models were launched in five AWS regions. These included US East (N. Virginia), Tokyo and Frankfurt. The image and multimodal-embedding models launched in two US regions.^[1]

The text embedding model had become generally available earlier. Titan Embeddings for text became generally available on 28 September 2023, the same day Bedrock itself became generally available.^[2] Its arrival before the text generation models meant the family's first generally available model was an embedding model, not a text generator.^[2]

Impact

The event filled out the Titan family with models for three purposes: generating text, creating images, and comparing text and images. Each came with a stated limit. Titan Text Lite has a smaller context window than Titan Text Express, the image model puts a watermark on each output, and the multimodal embedding model is aimed at short English text.^[1]

For readers, the launch shows how Amazon's first models were built out in stages, first with an embedding model, then with text and image models, all available through the same managed service.^[1][2]

Some terms in the launch may need explaining. An invisible watermark is a signal built into an image that software can detect but people normally cannot see, which helps show that the image came from the model. Inpainting fills in part of an existing picture, and outpainting extends a picture past its original edges. An embedding is a list of numbers that represents the meaning of a piece of content, and comparing embeddings is a common way to find related content.^[1]