Exploring Image Similarity through Generative Language Models: A Comparative Study of GPT-4 with Word Embeddings and Traditional Approaches

In this article, we propose a novel approach for determining image similarity, leveraging advancements in generative artificial intelligence. At the heart of our method is the use of OpenAI’s GPT-4 large language model for generating image captions, combined with the Ada v2 word embedding model for semantic analysis. This technique involves creating textual descriptions of images via GPT-4 and subsequently computing cosine similarity of these descriptions using Ada v2 word embeddings. We compare this innovative approach with traditional image similarity methods, with a particular focus on the VGG16 neural network approach, employing the DISC21 dataset for our analysis. Preliminary results demonstrate the promising potential of this method in the field of image similarity assessment. The paper delves into both the advantages and current limitations of our approach, including constraints like rate limits in experimentation and the rapidly evolving capabilities of language models in vision tasks. Our findings indicate a trajectory towards improved outcomes as these models continue to advance, underscoring the growing intersection of language and vision models in artificial intelligence for applications like image similarity evaluation.