Skip to content Skip to footer

Generating style item descriptions by refining a visual-language model using SageMaker and Amazon Bedrock.

Creating high-quality product descriptions for millions of products is a significant challenge in online retail. Machine learning (ML) and natural language processing (NLP) technologies can automate this task, vastly improving the searchability of products and increasing customer satisfaction. This article presents a method of predicting domain-specific product attributes from product images with a fine-tuned Vision-Language Model (VLM) on a fashion dataset using Amazon SageMaker and generating product descriptions from these predicted attributes using Amazon Bedrock.

VLMs have shown exemplary performance in tasks such as image captioning, text-guided image generation, and visual question-answering. VLMs can learn product attributes directly from images. For example, a model may identify a shirt as “long sleeve” and “cotton neck”, enhancing the searchability of products. This article uses BLIP-2, which consists of an image encoder, a Querying Transformer, and a Large Language Model, as the VLM.

The VLMs are trained on Amazon SageMaker, an ML service that manages necessary Amazon Elastic Compute Cloud (EC2) instances, provides the right Hugging Face container, uploads scripts, and downloads data from an S3 bucket to the container. The article also demonstrates how to deploy the fine-tuned VLM and predict product attributes using SageMaker.

The generated product attributes are then used to form product descriptions through Amazon Bedrock, a service offering high-performing foundational models from leading AI companies. These descriptions significantly contribute to customer satisfaction by personalizing the buying experience and improving recommendation algorithms, thereby increasing the likelihood of purchases.

This article concludes that leveraging VLMs on SageMaker and Large Language Models on Amazon Bedrock provides a powerful solution for automating fashion product description generation. This method can greatly benefit eCommerce platforms as it enhances both product searchability and the personalization of the platforms. This automation marks an efficient and ground-breaking development in online retail.

Leave a comment

0.0/5