---
license: apache-2.0
base_model: google/siglip2-base-patch16-224
tags:
  - onnx
  - vision
  - image-text-matching
  - nebula
---

# siglip2-base-patch16-224 (ONNX)

This is [Google's SigLIP 2 base/224](https://huggingface.co/google/siglip2-base-patch16-224) exported to ONNX format for CPU inference, used by [Nebula](https://github.com/diegohh0411/nebula) for local, offline image search.

## What's inside

| File | Description |
|---|---|
| `model.onnx` | Combined vision + text encoder (~110 MB) |
| `tokenizer.json` | SigLIP tokenizer |

## Model inputs & outputs

The single `model.onnx` file contains both encoders. You can run either independently by passing a dummy tensor for the unused branch.

**Inputs**

| Name | Shape | dtype |
|---|---|---|
| `pixel_values` | `[image_batch, 3, 224, 224]` | float32 |
| `input_ids` | `[text_batch, seq_len]` | int64 |

**Outputs**

| Name | Shape | dtype | Description |
|---|---|---|---|
| `image_embeds` | `[image_batch, 768]` | float32 | L2-normalizable image embedding |
| `text_embeds` | `[text_batch, 768]` | float32 | L2-normalizable text embedding |
| `logits_per_image` | `[image_batch, text_batch]` | float32 | Cosine similarity scores |
| `logits_per_text` | `[text_batch, image_batch]` | float32 | Cosine similarity scores (transposed) |

## How it was exported

```bash
optimum-cli export onnx \
  --model google/siglip2-base-patch16-224 \
  --task zero-shot-image-classification \
  --opset 18 \
  ./models/
```

Requires `optimum[onnxruntime]` and `transformers`.

## License

Inherits [Apache 2.0](https://huggingface.co/google/siglip2-base-patch16-224) from the original Google SigLIP 2 model.