siglip2-base-coreml

SigLIP 2 B/16-256 — Core ML

Core ML conversion of google/siglip2-base-patch16-256, split into separate image and text encoders for on-device text→image retrieval. Built for CuttyLA's footage search; usable by anything that wants SigLIP 2 on Apple silicon.

Files

Both encoders emit L2-normalized 768-d embeddings (embedding output); similarity is a plain dot product. Minimum deployment target: macOS 15.

Usage notes

• Image preprocessing is a squash-resize to 256×256 (no center crop), pixels scaled to [-1, 1]. The ImageType input already applies the scaling.
• Text must be tokenized with the bundled Gemma tokenizer and padded to 64 with the pad token (0), no attention mask — SigLIP was trained that way and embeddings drift if padding differs.
• Conversion is parity-gated: every release's embeddings match the PyTorch reference at cosine ≥ 0.99 on a fixture set. Conversion source: the models/siglip2 directory of the CuttyLA repository.

Versioning

Files in this repo are immutable once published. Re-conversions are published as new versions, never overwrites.

License

Apache 2.0, same as the original weights by Google. This repository redistributes a converted form of those weights without modification to their values beyond 8-bit palettization. The conversion is published by karaiman.team; the weights themselves are Google's work, not ours.