what should I use for image embeddings and similarity search?

For image embeddings + similarity search, a strong default is:

Best practical choice

Use a pretrained vision embedding model + a vector database/search index.

Good model options

• OpenCLIP / CLIP: very common, easy to use, strong baseline for semantic image search.
• SigLIP / SigLIP2: often better retrieval performance than CLIP.
• DINOv2: great for visual similarity, object/scene structure, less text-aligned than CLIP.
• ImageBind: if you want multimodal embeddings across image/audio/text, etc.

Good vector search options

• FAISS: great if you want an in-memory/local solution.
• Milvus / Weaviate / Qdrant: good for production vector search.
• Pinecone: managed option if you want less ops.

Which should you pick?

If you want:

• “Find images similar to this image”
  → DINOv2 or CLIP/SigLIP
• “Search images by text query”
  → CLIP or SigLIP
• Both text-to-image and image-to-image search
  → SigLIP or CLIP
• Highest visual similarity quality
  → DINOv2
• Best all-around semantic retrieval
  → SigLIP

Simple recommended setup

If you want a safe default:

• Embedding model: SigLIP or OpenCLIP
• Index/search: FAISS locally, or Qdrant/Milvus in production
• Distance metric: cosine similarity
• Preprocessing: resize/crop consistently, normalize embeddings, optionally store metadata

Rule of thumb

• Use CLIP/SigLIP if your search is semantic.
• Use DINOv2 if your search is about visual resemblance.
• Use a vector DB when you have more than a small number of images.

If you tell me your use case — e.g. product search, duplicate detection, art search, or multimodal search — I can recommend a specific stack and model size.