EmbeddingGemma2
An open-source multimodal embedding model released by Google, built on the Gemma 2 language model architecture. It encodes content from multiple modalities—including text, images, audio, and video—into unified 768-dimensional vectors, mapping content from different modalities into a shared vector space to support cross-modal semantic retrieval.
Core Facts
Timeline (last 90 days)
Google open-sourced the multimodal embedding model EmbeddingGemma2, which can convert text and images into the same vector space.
In addition to text and images, EmbeddingGemma2 also supports audio and video, converting content of different modalities into 768-dimensional vectors.
EmbeddingGemma2 has separate encoders for text and images, but outputs them into the same 768-dimensional shared vector space, so a sentence can be directly compared with an image.
EmbeddingGemma2 is based on the Gemma 2 language model architecture and was trained with multimodal contrastive learning.
Cross-modal retrieval results should not be filtered using a fixed numeric threshold, but should instead rely on ranking.
EmbeddingGemma2 employs Matryoshka Representation Learning, encoding the most important information at the front of the vector to support vector truncation.
Google states that searching with 256 dimensions can still retain approximately 95% of the quality.
After vector truncation, the length is no longer 1 (the length of the first 256 dimensions is approximately 0.709), so renormalization is required.
Truncating a 768-dimensional vector to 256 dimensions reduces data size by approximately three times (1 million photos shrink from approximately 3 GB to approximately 1 GB), and the top 9 search results are identical to those from the 768-dimensional version.
All Facts (9)
In addition to text and images, EmbeddingGemma2 also supports audio and video, converting content of different modalities into 768-dimensional vectors.
65%UnverifiedEmbeddingGemma2 has separate encoders for text and images, but outputs them into the same 768-dimensional shared vector space, so a sentence can be directly compared with an image.
50%UnverifiedEmbeddingGemma2 is based on the Gemma 2 language model architecture and was trained with multimodal contrastive learning.
50%UnverifiedCross-modal retrieval results should not be filtered using a fixed numeric threshold, but should instead rely on ranking.
50%UnverifiedEmbeddingGemma2 employs Matryoshka Representation Learning, encoding the most important information at the front of the vector to support vector truncation.
50%UnverifiedGoogle states that searching with 256 dimensions can still retain approximately 95% of the quality.
50%UnverifiedAfter vector truncation, the length is no longer 1 (the length of the first 256 dimensions is approximately 0.709), so renormalization is required.
50%UnverifiedTruncating a 768-dimensional vector to 256 dimensions reduces data size by approximately three times (1 million photos shrink from approximately 3 GB to approximately 1 GB), and the top 9 search results are identical to those from the 768-dimensional version.
50%UnverifiedGoogle open-sourced the multimodal embedding model EmbeddingGemma2, which can convert text and images into the same vector space.
60%