Pure Rust image understanding engine based on SigLIP 2. Zero-shot image classification, image embeddings, and image-text similarity. No Python runtime, no CUDA, no external dependencies.

score = sigmoid(cosine_sim(image_embed, text_embed) * exp(logit_scale) + logit_bias)

Image (224x224) → Patch Embedding (196 patches)
    → Add Position Embeddings
    → 12x ViT Transformer Layers (bidirectional)
    → Post-LayerNorm
    → MAP Pooling (cross-attention with learned probe)
    → L2 Normalize
    → 768-dim Image Embedding

Text → Tokenize → Token + Position Embedding
    → 12x ViT Transformer Layers
    → Final LayerNorm (last token)
    → Linear Head
    → L2 Normalize
    → 768-dim Text Embedding

Score = sigmoid(cosine_sim * exp(scale) + bias)

siglip-model/
  qora-vision.exe      - 4.4 MB    Inference engine
  model.qora-vision    - 210 MB    Full model (vision + text, Q4)
  tokenizer.json       - 33 MB     Text tokenizer (256K vocab)
  config.json          - 611 B     QORA-branded config
  README.md            - This file

# Zero-shot classification (fast, from binary)
qora-vision.exe siglip --load model.qora-vision --image photo.jpg --labels "cat,dog,bird,car"

# Image-text similarity
qora-vision.exe siglip --load model.qora-vision --image photo.jpg --text "a photo of a sunset"

# Image embedding only
qora-vision.exe siglip --load model.qora-vision --image photo.jpg

SigLIP 2 improves over the original SigLIP with enhanced semantic understanding, localization, and dense features. The sigmoid loss enables better calibrated scores compared to CLIP's softmax-based approach.

All tests run with Q4 quantization on CPU.

Input: Solid red 224x224 image Labels: red, blue, green, yellow

Input: Solid blue 224x224 image Labels: red, blue, green, yellow

Input: Solid green 224x224 image Labels: "a photo of a cat", "a photo of a dog", "a solid green image", "a landscape"

cargo build --release                       # CPU
cargo build --release --features gpu        # GPU (Windows/Linux, Vulkan)
cargo build --release --features gpu-metal  # GPU (macOS, Metal)

• cortex — Rust deep learning framework (GPU via wgpu/Vulkan/Metal)
• half — F16 support
• image — Image loading (PNG/JPEG)
• tokenizers — HuggingFace tokenizer
• safetensors — Weight loading
• serde_json — Config parsing

Built with QORA - Pure Rust AI Inference