Convolution, the operation underneath blur/sharpen/edge detection, made tangible: nine weights, one live preview. Edit the 3×3 kernel and the GPU re-filters the image as you type; presets (identity, box blur, sharpen, Sobel, emboss) show the classics.
Drop any image or start from the generated sample — the pixels are uploaded to a GPU texture and never leave the page; export the result as PNG.
Frequently asked questions
Why does the Sobel preset look gray and washed out?
Edge detection produces signed values around zero; the +0.5 offset maps them into the visible range. Negative weights are not just allowed — they are the whole point of edge kernels.
What is the divisor for?
It normalizes the kernel sum: 9 ones divided by 9 is a box blur (brightness preserved); divide by 1 for sharpen where the sum is already 1.
中文说明
卷积——模糊/锐化/边缘检测背后的那一个运算——变得可触摸：九个权重，一块实时预览。编辑 3×3 卷积核，GPU 随输入即时重滤；预设（恒等、盒式模糊、锐化、Sobel、浮雕）展示经典组合。
拖入任意图片或从内置样例开始——像素只上传到 GPU 纹理，从不离开页面；结果可导出 PNG。
常见问题
为什么 Sobel 预设看起来发灰？
边缘检测产生围绕零的有符号值；+0.5 偏移把它们搬进可见区间。负权重不仅允许——那正是边缘核的全部要点。
除数是干什么的？
归一化核内权重之和：九个 1 除以 9 是盒式模糊（保持亮度）；锐化的和本来就是 1，除以 1 即可。