Embed Images with DINOv3

DINOv3 is Meta's self-supervised vision encoder — one of the strongest open image-embedding models available. Tenzro's vision runtime serves DINOv3 (vits16 / vitb16 / vitl16), DINOv2, CLIP, and SigLIP2 (base / large / so400m). This tutorial downloads dinov3-vitb16, embeds an image, and compares two images by cosine similarity.

1. Download with license acceptance

# DINOv3 ships under Meta's CommercialCustom license — explicit accept required
tenzro model download dinov3-vitb16 --accept-license dinov3

# Output:
# License gate: dinov3 (CommercialCustom)
#   Terms: https://ai.meta.com/resources/models-and-libraries/dinov3-license
#   Acknowledgment: recorded in CF_MODELS
# Resolving artifact bundle from HuggingFace Hub...
#   Source: facebook/dinov3-vitb16 (ONNX export)
#   Files: model.onnx
#   SHA-256 verified
#   Saved to: ~/.tenzro/models/dinov3-vitb16/

2. Load into the vision runtime

# Load into the vision runtime
tenzro vision load dinov3-vitb16

# Output:
# Vision runtime loaded:
#   Model: dinov3-vitb16
#   Modality: vision
#   Output dim: 768
#   Normalization: ImageNet (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
#   L2-normalized: true

3. Embed an image

The runtime decodes PNG/JPEG/WebP via the imagecrate, Lanczos3-resizes to the encoder's expected resolution, applies ImageNet normalization, and runs the ONNX session. The output is L2-normalized so cosine similarity reduces to a dot product.

# Embed a single image (PNG, JPEG, or WebP — auto-detected)
tenzro embed-image \
  --model dinov3-vitb16 \
  --image ./photo.jpg

# Output:
# Embedding (dim=768, L2-normalized):
#   [0.0142, -0.0731, 0.0418, ..., 0.0294]
#   latency_ms: 38

4. Compare two images

# Compute cosine similarity between two images directly
tenzro vision similarity \
  --model dinov3-vitb16 \
  --image-a ./cat1.jpg \
  --image-b ./cat2.jpg

# Output:
# cosine_similarity: 0.847
# latency_ms: 71

5. Embed via JSON-RPC

# Equivalent JSON-RPC call. Image is base64-encoded raw bytes.
IMAGE_B64=$(base64 -i ./photo.jpg)

curl https://rpc.tenzro.network \
  -X POST \
  -H "Content-Type: application/json" \
  -d "{
    \"jsonrpc\": \"2.0\",
    \"id\": 1,
    \"method\": \"tenzro_visionEmbed\",
    \"params\": {
      \"model_id\": \"dinov3-vitb16\",
      \"image_base64\": \"$IMAGE_B64\"
    }
  }" | jq

A typical response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "model_id": "dinov3-vitb16",
    "embedding": [0.0142, -0.0731, 0.0418, "..."],
    "dim": 768,
    "l2_normalized": true,
    "latency_ms": 38
  }
}

See also

• Model serving documentation — license tiers, modality routing, the vision catalog
• Inference RPC reference—tenzro_visionEmbed, tenzro_visionSimilarity, tenzro_visionClassify
• Embed text with Qwen3-Embedding — pair text embeddings with image embeddings for retrieval