What to try
- Switch between Original, Box blur, Sharpen and Find edges. Try the rings, checkerboard and wave patterns.
- Click a pixel or move with the arrow keys, edit a weight or the divisor, then follow the neighborhood, rotated weights and output brightness below. Step or animate the scan to move across the image.
What to look for: Blur shares brightness with the neighbors. Sharpen strengthens local differences. An edge kernel cancels flat regions; positive responses light up while negative results are clipped to black.
How it works
Multiply each neighborhood value by the corresponding coefficient and sum the products. A normalized all-ones 3 × 3 matrix averages nine samples. Negative weights can subtract neighbors, making edges and directional differences visible. A bias shifts the result before it is clamped to a displayable channel value.
The playground shows the original and filtered images, then breaks down the pixel you select. It rotates the kernel 180° for mathematical convolution and repeats the closest sample at image boundaries. The Lab instead applies the displayed matrix directly (correlation). The symmetric presets behave identically either way; rotate an asymmetric matrix to reproduce this study in the Lab.
Use it in the Lab
Use Custom kernel in the Lab for 3 × 3 through 9 × 9 matrices, channels, bias, boundaries and repeated passes. Save a named kernel once its effect is useful and record it in a chain.
Limits & distinctions
RGB-only kernel processing keeps alpha unchanged. For transparency-aware softening, compare the dedicated blur tools, which handle premultiplied color during filtering.
Further reading
The Sliding Window
A convolution kernel (or filter) is a small matrix — typically 3×3 or 5×5 — that is slid across every pixel of an image. At each position, the kernel values are multiplied element-wise with the underlying pixel values and summed to produce a single output pixel.
// 3×3 Sharpen kernel
const sharpen = [
[ 0, -1, 0],
[-1, 5, -1],
[ 0, -1, 0]
];
// Apply: for each pixel (x, y)
let sum = 0;
for (let ky = -1; ky <= 1; ky++)
for (let kx = -1; kx <= 1; kx++)
sum += image[y+ky][x+kx] * sharpen[ky+1][kx+1];Common Kernels
Kernel Zoo
Gaussian blur smooths noise. Sobel detects horizontal/vertical edges. Laplacian finds contours. Unsharp mask enhances detail. Emboss creates a 3D relief effect. Each is just a different set of matrix weights.
Different kernels produce dramatically different effects. The Gaussian blur kernel averages nearby pixels weighted by a bell curve. The Sobel operator uses two kernels — one for horizontal edges, one for vertical — and combines them via √(Gx² + Gy²).
Chaining Operations
In BVDART, convolution is the namesake technique. Artworks are built by composing chains of kernel operations — blur, then edge-detect, then sharpen — each one revealing hidden structure in noise fields and procedural textures. The order matters: the same kernels in a different sequence produce entirely different compositions.
Try chaining a Gaussian blur (σ=2) followed by a Laplacian edge detect — this is the Laplacian of Gaussian (LoG) operator, excellent for finding edges at a specific scale.
Articles
Related artworks
Related Lab categories
Sources & references
Guide updated