Add SigLIP 2 B/16-256 Core ML encoders and tokenizer

ImageEncoder.mlpackage.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:426115f240ead5faf69b073e08dd1b959d850ca5c592537cd81886992283b2fb
+size 91700398

README.md

@@ -0,0 +1,49 @@
+---
+license: apache-2.0
+base_model: google/siglip2-base-patch16-256
+tags:
+  - coreml
+  - siglip2
+  - image-text-retrieval
+library_name: coreml
+---
+
+# SigLIP 2 B/16-256 — Core ML
+
+Core ML conversion of [google/siglip2-base-patch16-256](https://huggingface.co/google/siglip2-base-patch16-256),
+split into separate image and text encoders for on-device text→image retrieval.
+Used by the Resource Library in ScreenKite and PilotCut.
+
+Redistributed from [palmier-io/siglip2-base-coreml](https://huggingface.co/palmier-io/siglip2-base-coreml)
+under Apache 2.0. The files are byte-identical to that repository.
+
+## Files
+
+| File | Contents |
+|---|---|
+| `ImageEncoder.mlpackage.zip` | Vision tower, 256×256 input, 8-bit palettized (per-grouped-channel) |
+| `TextEncoder.mlpackage.zip` | Text tower, 64-token input, 8-bit palettized |
+| `tokenizer.zip` | Gemma SentencePiece tokenizer files (`tokenizer.json`, config) |
+| `manifest.json` | File names, sha256s, sizes, model dims |
+
+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.
+
+## 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.

TextEncoder.mlpackage.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48f80e35ce40a9dcdc55bef986a104d3153e1cfa78229bb45c4724f3f3427368
+size 258593083

manifest.json

@@ -0,0 +1,25 @@
+{
+  "model": "siglip2-base-patch16-256",
+  "version": 1,
+  "weights": "palettized-8bit",
+  "embeddingDim": 768,
+  "imageSize": 256,
+  "contextLength": 64,
+  "files": {
+    "imageEncoder": {
+      "name": "ImageEncoder.mlpackage.zip",
+      "sha256": "426115f240ead5faf69b073e08dd1b959d850ca5c592537cd81886992283b2fb",
+      "bytes": 91700398
+    },
+    "textEncoder": {
+      "name": "TextEncoder.mlpackage.zip",
+      "sha256": "48f80e35ce40a9dcdc55bef986a104d3153e1cfa78229bb45c4724f3f3427368",
+      "bytes": 258593083
+    },
+    "tokenizer": {
+      "name": "tokenizer.zip",
+      "sha256": "c37f2a8e8555d8561109564c4f60ee962b0072abddcfcfd599d321469d6d1ef5",
+      "bytes": 5460173
+    }
+  }
+}

tokenizer.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c37f2a8e8555d8561109564c4f60ee962b0072abddcfcfd599d321469d6d1ef5
+size 5460173