Loading AI models
This may take a moment on first visit
Everything runs on your device. No images are uploaded anywhere.
When you look at a photograph, you see a face, a landscape, a moment. A machine sees none of that. It sees a grid of numbers — millions of tiny colored squares called pixels, each defined by three values: Red, Green, and Blue (RGB).
Computer vision is the process of turning that grid of numbers into meaning: "this is a dog," "this person looks happy," "this is a stop sign." Upload a photo to see how.
AI models don't just see — they describe. But their words reveal more about their training data than about your photo. Upload an image to hear the machine narrate what it "sees."
A surveillance camera doesn't see a person. It sees data points. Choose a source and toggle detection layers to see what gets extracted.
Face Analysis — BlazeFace detects faces and landmarks; face-api.js (SSD MobileNet v1 + AgeGenderNet + FaceExpressionNet) estimates age, gender, and expression. A crude BMI estimate is derived from the face bounding box width-to-height ratio — research (Coetzee et al., 2009) shows facial WHR loosely correlates with body mass, but the estimate is intentionally unreliable, demonstrating how surveillance systems extract body data from minimal facial geometry.
Pose — Facial geometry derived from BlazeFace's 6 keypoints (eyes, nose, mouth, ears).
Objects — COCO-SSD identifies 80 everyday object categories (person, car, dog, chair, etc.).
Content — NudeNet detects and censors exposed body parts (16 categories). Black boxes are drawn over NSFW regions.
Scene — MobileNet v2 classifies the image against 1,000 ImageNet categories.
All models run locally in your browser via TensorFlow.js. No data is sent to any server.
Machine vision is fragile. Upload any photo and see what the model thinks. Then try to fool it.
Beauty filters are not magic. They are a series of deliberate geometric decisions about what a face should look like. Upload a selfie to see the machine reshape you.
Face recognition doesn't see faces — it sees numbers. Every face is compressed into a list of 128 numbers called a face descriptor. Two faces are "the same person" if their numbers are close enough. Upload two photos to see how the machine measures identity.