mtbui2010/siglip-base-patch16-224-ONNX

SigLIP base-patch16-224 — image and text towers (ONNX)

ONNX exports of both towers of google/siglip-base-patch16-224 (Apache-2.0), as served by VisionServe — crop rescoring in the rfdetr-gdino-siglip* open-vocabulary router.

visionserve pull siglip-image
visionserve pull siglip-text
visionserve pull rfdetr-gdino-siglip       # pulls both towers as dependencies

Files

file	contents
image/model.onnx + image/model.onnx.data	vision tower, opset 17, fp32 (external weights; keep the pair together)
text/model.onnx + text/model.onnx.data	text tower, opset 17, fp32
text/tokenizer.json	the checkpoint's SentencePiece Unigram tokenizer (32 000 pieces)

I/O contract

tower	input	output
image	pixel_values f32 [N, 3, 224, 224]	image_embeds f32 [N, 768], not L2-normalised
text	input_ids i64 [N, 64]	text_embeds f32 [N, 768], not L2-normalised

• Image preprocessing: resize to exactly 224×224 (squash, bicubic, no centre crop), /255, mean = std = 0.5 per channel → [-1, 1]. CLIP's constants are wrong here and fail silently.
• Text: pad to 64 tokens with </s> (id 1), not <pad> (id 0). SigLIP pools with attention and reads the padding; padding with id 0 moves embeddings 0.706 cosine away from the reference.

Both exports were verified against the PyTorch checkpoint before being written (image: 32 real crops; text: per-prompt cosine floor 0.9999).

License

Apache-2.0, inherited from google/siglip-base-patch16-224.