Extract Image Embedding (OpenCLIP)

@laion1 VersionExtract

Passes each Image through an OpenCLIP image tower — the architecture arch built to match the weights bound to model — and emits one Embedding, optionally L2-normalised for cosine or inner-product retrieval. Use for image-side retrieval, clustering, or similarity; pair with a matching CLIP text encoder for cross-modal search.

How it fits

Typical backends

Embed frames from files and forward the vectors over HTTP for ingestion into a vector index.

Pull frames over HTTP, segment regions, crop them, and embed each crop for region-level visual search.

Caveats

Hard constraintThe image tower is built from a workspace-owned weights file bound into the `model` slot, loaded offline at startup into the open_clip architecture named by arch; no HuggingFace download happens at deploy. arch MUST match the bound checkpoint (a mismatch fails at weight load), and the slot must be bound or the deploy aborts.
Hard constraintOutput dimensionality is set by the bound arch/weights (e.g. 1152 for SigLIP2 SO400M-16, 768 for ViT-L-14, 1024 for ViT-H-14), not a free parameter; binding a different-width checkpoint yields vectors an index or text tower built on another width will reject.
Fallbackdevice requesting CUDA resolves to CPU when no compatible accelerator is visible at startup; inference continues at CPU latency.
Parameter interactionnormalize produces unit-length vectors so cosine similarity reduces to a dot product. Disable it only when the consumer wants raw, un-normalised features.
LimitationThis is the image tower only; it carries no text encoder. Cross-modal retrieval requires a separate text encoder using the identical arch and weights.
CompatibilityInputs are converted to RGB internally before preprocessing, so BGR, grayscale, or RGBA upstream frames are handled without extra preparation.
State lifetimeThe bound model, arch, device, and normalisation are read once at startup; changing the slot binding, architecture, device, or normalisation requires a redeploy.

Versions

38782d1elatestdefault
Automated release
6/29/2026