SigLIP 2 Large, Core ML

Core ML conversion of Google's google/siglip2-large-patch16-256, for on-device image search in the Protos Mac app.

Changes from the original: the model is split into an image tower and a text tower, each exported to Core ML in float16 with coremltools 8.3.0. The image tower takes pixel_values float32 [1, 3, 256, 256] scaled to -1..1 and returns image_embeds; the text tower takes input_ids int32 [1, 64] (Gemma tokens) and returns text_embeds. Both return the raw projected vectors; L2-normalise them yourself. The weights are otherwise unchanged. The tokenizer files are Google's, copied as they are.

License: Apache-2.0, as the original. Copyright Google LLC; this conversion is distributed under the same license, with the changes stated above.

Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for protos-photos/siglip2-large-patch16-256-coreml

Quantized
(4)
this model