Google's Gemini Embedding 2: A Delicious Dive into Image Similarity
research#embeddings📝 Blog|Analyzed: Mar 22, 2026 02:45•
Published: Mar 22, 2026 02:44
•1 min read
•Qiita VisionAnalysis
Google's Gemini Embedding 2 is making waves by enabling the analysis of image similarities across diverse datasets. The article highlights an exciting experiment using different curry images, demonstrating the model's ability to understand visual nuances. This showcases a significant step forward in how we can leverage the power of embeddings for image recognition.
Key Takeaways
- •Google's Gemini Embedding 2 is a new model that can map text, images, videos, audio, and PDFs into a single embedding space.
- •The experiment focuses on understanding image similarities using various curry images.
- •The process involves converting images into embeddings, comparing them using cosine similarity, and using an LLM to explain the similarities.
Reference / Citation
View Original"To check “how much can it be used with just images?”, I embedded images of 8 kinds of curry dishes and compared them by cosine similarity, and then I constructed a pipeline to make Gemini 2.5 Flash explain “why they are similar”."