Creative Technology · Computer Vision
AI-Powered
Image Segmentation
Four ways to answer the same question: which pixels belong to what? Start with a hand-written threshold rule, then hand the job to three pretrained models.
Choose an image
Start with a tested example, or try a photo of your own.
Start here · COCO examples
Four images are drawn at random from 24 bundled COCO validation examples.
Useful for seeing where a model trained on one dataset stops making sense.
Choose a method
Same image, four different ideas about what a segment is.
Pixel brightness distribution. The marker shows where the threshold falls.
There is nothing to tune. SegFormer runs at a fixed input size and returns its best guess per pixel.
Lower values keep more objects, including ones the model is guessing at.
Run once to compute the image embedding, then click anywhere on the picture to segment the object under your cursor. Each click is independent — try a few.
Choose a COCO example or upload your own image.
Result
Run a method to see its segments
Choose an image first. The original image appears here before you run anything.
Click the image to segment the object under your cursor.
| Segment | Detail |
|---|---|
| Run a method to see its segments. | |
Compare
The same image, four ways
Every run is kept here, so you can put the four methods side by side without re-running them.
Run a method and its result will appear here.
What you are doing
Segmentation asks a harder question than detection. Not "is there a cat?" but "exactly which pixels are the cat?"
A detector draws a box around an object. A segmenter has to decide, pixel by pixel, where that object ends — which is why the four methods here disagree so much. Thresholding uses one hand-chosen rule about brightness. SegFormer labels every pixel with a scene class but cannot tell two people apart. MaskFormer predicts one mask per object. SAM takes a click as its question and returns whatever object sits under that point, without any fixed class list.
Try this: run Threshold on an image with uneven lighting, then run it again on a clean studio-like scene. The rule does not change — only the picture does. That gap is the reason the other three methods exist.