Amazon has introduced the Titan Image Generator G1, a cutting-edge text-to-image model available through Amazon Bedrock. This model can interpret prompts describing multiple objects within different contexts and reflects these details in the images generated. It also provides advanced image editing options such as smart cropping, in-painting, and background changes. Currently, the service is available in US East (N. Virginia) and US West (Oregon) AWS Regions.
Titan Image Generator has broad usage across multiple industries, including architecture, manufacturing, fashion design, and game design. It can boost employee creativity by allowing them to visualize designs simply through textual descriptions. It also enhances customer experience through personalized advertising and visual chatbots.
Users can further fine-tune the Titan Image Generator using their own data through custom models for Amazon Bedrock. This allows them to create images with unique characteristics in custom data sets that fit their specific requirements, such as brand guidelines or specific campaign styles.
An example provided in the post talks about fine-tuning the Titan model to recognize two new categories, ‘Ron the dog’ and ‘Smila the cat’. The process involves preparing the data for model fine-tuning, creating a model customization job through Amazon Bedrock, and finally deploying the fine-tuned model using Provisioned Throughput.
Data privacy is ensured with fine-tuning data and custom models remaining private in the user’s AWS account. They are not used for model training, service improvements, nor shared with third-party model providers.
The post also explains the process of data preparation for a customization job, with images and captions needed in a JSONL format with multiple JSON lines. A demonstration using Ron and Smila pictures provides a user-friendly guide on how to use the platform. The Titan Image Generator G1 can handle fine-tuning jobs with 5 to 10,000 images, and better accuracy is achieved with more variations of the style or class.
After the training data has been prepared, Amazon Bedrock can be used to embark on a new customization job. This involves selecting custom models, providing fine-tuned model names, job configurations, input data, output data, and IAM role permissions.
The fine-tuned models are then deployed via Provisioned Throughput for consistent performance. Users can purchase Provisioned Throughput, set model units and commitment term, and start testing their models immediately.
Using fine-tuning to customize the model for Ron and Smila resulted in the best hyperparameters being 5,000 steps with a batch size of 8 and a learning rate of 1e-5. Examples showcasing the successful results are also shared in the post.
The authors suggest users can adapt the example provided to generate hyper-personalized images using Amazon Titan Image Generator for their specific use case.
