Kernel Lab
Nine numbers are enough to blur, sharpen, or find every edge in a photograph. Edit them below and watch a 3×3 convolution rewrite a real photo live, on every keystroke — a cat, a rocket launch, a Beijing temple, or any picture you upload from your own machine. Then drag a window over one pixel to see the nine multiply-adds that produced it, or hit Scan to watch the whole output painted a dot-product at a time.
Photo: Stefan van der Walt, CC0 (scikit-image “chelsea”) · shown in grayscale — a real CNN slides the same kernel over each color channel
Edge (Laplacian) — responds wherever brightness changes in any direction — outlines everything at once.
One output pixel, from the inside
9 pixels × 9 weights = 9 signed products → Σ → one number — the whole convolution, atom by atomthis is the exact number written into the highlighted output pixel: Σ of the nine products, shown as |value|. Scan repeats it for all 25,600 pixels in raster order.
A CNN doesn’t use hand-written kernels — it learns them. See the CNN X-Ray playground for the learned versions of these filters.
160×160 convolution recomputed on every keystroke — this is the exact operation inside every CNN layer. Uploads accept any image your browser can display (PNG, JPEG, BMP, WebP, GIF…) and stay on your machine.