L3N0X/siglip2-base-patch16-224-litert

SigLIP 2 base (patch16, 224) vision tower, LiteRT fp16, inline weights

A repackaged copy of litert-community/SigLIP2-base-patch16-224's siglip2_base_224_fp16.tflite (commit 509b5cbcf1a849f37696be08f8297c6cd3050bf4), used by the EatMe app to name loose food from a photo, entirely on the device.

What changed

Only the file layout. The original stores 209 of its 240 constant buffers after the flatbuffer, referenced by Buffer.offset/size. The TFLite interpreter only resolves those when it is built with the model's allocation, and the LiteRT C API (TfLiteInterpreterCreate, used by e.g. react-native-fast-tflite on Android and iOS) builds it without one: the weights load as empty tensors and the first Invoke fails with Input tensor 177 lacks data.

This copy stores every buffer inline, 16-byte aligned as the schema asks. The tensors, ops and bytes of every weight are unchanged; outputs are bit-identical to the original.

File	siglip2_base_224_fp16.tflite
Size	185,436,176 bytes
SHA-256	f4c05fa62edfabeba8d72bb261db274abd8cd2395d3733db6ad4b72fba655928
Input	float32[1, 3, 224, 224], RGB scaled to −1…1
Output	image embedding (normalise before cosine similarity)

License

Apache 2.0, as the original. SigLIP 2 is by Google; the LiteRT export is by the litert-community organisation.