Static gallery
Safe PNG assets, captions, tags, and metadata are served directly by GitHub Pages.
Browser-based multimodal AI · GitHub Pages
Search images with natural language using CLIP
A static, backend-free portfolio application for image-text retrieval and zero-shot image classification using quantized ONNX inference in the browser.
Initializing static gallery…Transformers.js + ONNX Runtime Web
—safe gallery images
—visual categories
—embedding dimensions
0backend servers
Interactive demo
One CLIP model, two multimodal workflows
Try:
ModeGallery preview
Query latency—
MetricCosine similarity
PNG, JPEG, or WebPProcessed locally in your browser · maximum 10 MB
PredictionUpload an image to begin
Latency—
| Rank | Label and prompt | Similarity | Relative score |
|---|---|---|---|
| No predictions yet. | |||
Relative scores are softmax-normalized only across the candidate labels shown. They are not calibrated confidence values.
Architecture
Static hosting with real browser inference
CLIP encoders
The text and vision encoders load as quantized ONNX models through Transformers.js while the site itself is published from main/docs.
Shared embeddings
Images and natural-language queries are represented in the same 512-dimensional space.
Client-side ranking
Cosine similarity, top-k ranking, latency, and UI rendering all run in JavaScript.
Evaluation framework
Metrics are measured, never invented
Correct image appears first.
Correct image appears in top five.
Correct image appears in top ten.
Python CPU text-encoding benchmark; browser latency is shown live above.
Measured on 10 curated retrieval queries across a 12-image demonstration gallery. Results apply only to this small public evaluation set; browser latency is still measured live.
Model details
CLIP ViT-B/32
CLIP learns aligned image and text representations. Higher cosine similarity generally indicates stronger alignment, but does not guarantee semantic correctness.
Limitations
Human review remains necessary
Prompt wording, gallery coverage, quantization, browser resources, and model bias can affect results. This demo is not suitable for consequential decisions.