IMAGE–TEXT SIMILARITY · RUNS ON YOUR DEVICE
Image–Text Similarity
Type several captions and see how well each one matches an image. MobileCLIP-S0 maps text and image into one embedding space; the cosine similarity is the match score.
Powered by Apple MobileCLIP
MOBILECLIP-S0 · IN-BROWSER
Checking this browser…
JPEG, PNG or WebP · up to 10 MB · Ctrl/Cmd+V to paste
Cosine similarity is in roughly the 0–0.4 range for CLIP-style models; compare rows, not absolute values.
Runs the Transformers.js port of Apple MobileCLIP-S0 (first generation). MobileCLIP2 is not yet available for the browser; use the Python guide for S0–S4 of the second generation.
Match scores
Results and measured timings will appear here.
Images are resized to 512 px on the longest side. Scores describe similarity, not truth—test unrelated labels too.
Images, labels and results stay on your device. Usage events record the model, step and timings only.
Questions
More toolsWhat does zero-shot mean?
The model was never trained on your labels. It compares the image embedding with a text embedding of each label, so any phrase can be a class.
How do I write good labels?
Use short, concrete noun phrases and keep the template on. Add a catch-all like “something else” to see whether the model is unsure.
Is this MobileCLIP2?
No. The browser tool uses MobileCLIP-S0, the first-generation model ported to Transformers.js. MobileCLIP2 runs locally with Python; see the guide.
Is my image uploaded?
No. Embedding happens in a Web Worker in your browser. Only the model files are downloaded from Hugging Face on first use.