Embeddings
DINOv3dinov3
Image embeddings for similarity and retrieval. Built with DINOv3
0.05 per image
Public
Embeddings
SigLIP 2 API: image and text embeddings
SigLIP 2 embeds images and texts in one vector space, so images can be matched to text queries and labels without training.
Price
| Billed item | Unit | Credits |
|---|---|---|
| Image embedding One image per request. | per image | 0.05 ≈ US$0.00005 top-up |
| Text embedding Each string in texts[]. | per text input | 0.01 ≈ US$0.00001 top-up |
Included in the public pool: 100 calls per day after sign-up. Prices are in credits. Failed calls cost nothing.
Sample results
Recorded responses of the API on sample images.
- a dog on a beach0.133
- a dog in the snow0.076
- a beach at sunset0.040
- a bicycle0.018
- a cat on a sofa0.000
- a bowl of ramen0.133
- 一碗拉面0.108
- a cup of coffee0.055
- a salad0.035
- a pizza0.017
- a rainy street at night0.154
- neon signs0.072
- a snowy road0.054
- a sunny park0.014
Request and response
POST https://api.stack.foldbun.com/v1/embeddings with a Bearer API key. Each call carries either one image or a list of texts.
Request
# macOS: replace "base64 -w0 beach-dog.webp" with "base64 -i beach-dog.webp" curl https://api.stack.foldbun.com/v1/embeddings \ -H "Authorization: Bearer $FOLDBUN_API_KEY" \ -H "Content-Type: application/json" \ -d @- <<EOF {"model": "siglip2", "image": {"base64": "$(base64 -w0 beach-dog.webp)"}} EOF
import base64, os, requests with open("beach-dog.webp", "rb") as f: image = base64.b64encode(f.read()).decode() r = requests.post( "https://api.stack.foldbun.com/v1/embeddings", headers={"Authorization": f"Bearer {os.environ['FOLDBUN_API_KEY']}"}, json={"model": "siglip2", "image": {"base64": image}}, timeout=60, ) r.raise_for_status() print(r.json()) # vectors are L2-normalised: similarity = dot product
Response (shortened)
{
"id": "call_01a1913e-bed1-7605-b1be-55da62598752",
"model": "siglip2",
"image": {"height": 1008, "id": "img_d94c2292…17f8", "width": 1512},
"dimensions": 768,
"embeddings": [
{"index": 0, "vector": [0.0067, -0.0157, -0.0138, 0.0029, …]}
],
"cost": 0.05,
"balance": 1234.51
}Request fields
| Field | Description |
|---|---|
| model | siglip2. |
| image | {"base64": "…"} or {"id": "img_…"} from an earlier response. Exactly one of image and texts. |
| texts | 1–64 strings of 1–1000 characters each. |
Pricing examples
| Call | Credits |
|---|---|
| 1 image 0.05 | 0.05 ≈ US$0.00005 top-up |
| 10 texts 10 × 0.01 | 0.1 ≈ US$0.0001 top-up |
| 1,000 images 1,000 × 0.05 | 50 ≈ US$0.05 top-up |
| 1,000 texts 1,000 × 0.01 | 10 ≈ US$0.01 top-up |
Limits
- Images
- JPEG, PNG, static WebP; up to 20 MiB and 40 megapixels
- Coordinates
- Pixels of the original image with EXIF orientation applied; masks can be smaller and carry their own size
- Texts per call
- 1–64, each 1–1000 characters
- Vector
- 768 dimensions, L2-normalised float32
- Rate limit
- Depends on the account tier; see
GET /v1/models
Frequently asked questions
How is SigLIP 2 billed?
Images cost 0.05 credits each and texts 0.01 credits each. The model is in the public pool: 100 calls per day at no charge after sign-up. Failed calls cost nothing.
How are images matched to texts?
Image and text vectors share one space and are L2-normalised, so their dot product is the cosine similarity. Ranking candidate texts by that value gives zero-shot labels for an image.
Can an image and texts be sent together?
No. Each call carries exactly one of image and texts. Two calls, one for the image and one for the texts, produce vectors that can be compared.
Related models
Foldbun Stack — pay-per-use vision model API
Call SigLIP 2 through the REST API; the docs list request and response formats.
Read the docsGet an API keyComing soon