Client-side inference · zero backend

Vision Lab

Five multimodal AI capabilities running entirely inside this browser tab via transformers.js — no API calls, no server, nothing you upload ever leaves your machine.

backenddetecting…
models cached0 / 4
statusidle
Drop an image, click to browse, or paste from clipboard
JPG · PNG · WEBP

Image Captioning

Generates a natural-language description of the uploaded image using an encoder–decoder vision-to-text model.

model: Xenova/vit-gpt2-image-captioning
Generated caption

Zero-Shot Object Detection

Define any set of objects to look for — the model locates and draws bounding boxes around them in the image above, even for categories it wasn't explicitly trained to detect.

model: Xenova/owlvit-base-patch32
Detections

Zero-Shot Classification

Define any set of candidate labels — the model has never been trained on them specifically, yet ranks them by visual-semantic similarity via CLIP's shared embedding space.

model: Xenova/clip-vit-base-patch32
Ranked labels

Optical Character Recognition

Extracts printed text from the image using a transformer-based OCR model. Works best on clear, printed (not handwritten) text.

model: Xenova/trocr-small-printed
Extracted text

Semantic Image Search

Upload a batch of images to build a gallery, then search it with a natural-language query. Each image is scored against your query in CLIP's shared embedding space and ranked by relevance — no keywords or tags required.

model: Xenova/clip-vit-base-patch32 (shared with Classify)
Drop multiple images to build a searchable gallery
System log