We’ve been adding CUDA support to many of our custom tools lately and finding speed-ups of 40x in some cases. This tool is an example of an idea that would be totally impractical (hence the krazy) without a highly parallel GPU-based approach. The algorithm takes each pixel in the input image, and finds the shortest distance (in RGB space) from it to the pixels in the “selection” image. This requires an every-pixel-to-every-pixel distance calculation, with the final shortest distance value compared to a threshold to determine whether or not to mask the pixel.

Download KrazyKey 1.0

A keying operation masks out a subset of the RGB color cube. This subset is usually a contiguous region within the cube, with the final masking weights related in some way to the parameters of the shape’s geometry.

As an example, the ChromaKeyer makes an octahedron, but not a regular one; think of it as a box aligned to the color cube, but with 2 corners truncated along the white/black “desaturated” axis. Primatte uses a a regular polyhedron but with the vertices repositioned to enclose the appropriate bounding region.  The main algorithm uses 128 faces in a closed shape. In both these cases, forcing the sampling of a contiguous region may include color coordinates that you don’t want, no matter how complex the shape within the RGB cube.

KrazyKey does not create a polyhedron at all, but uses a point cloud to define the sampling region.  It can thus enclose ANY color set at all, regardless of shape, and does not have to be contiguous.  This means that any keying operation can be performed with just one KrazyKey, whereas you might need a few Primattes, or dozens of ChromaKeyers.

Chad put together a demo comp that shows the subset of the color cube used by several common keying methods compared to what’s possible with KrazyKey.

Comparing keying geometries (red/cyan)

KrazyKey’s interface is pretty simple:

You input two images, the first (Input) is the image you want the key applied to, and the second (Selection) is the set of colors you want to compare against.  Any pixels with an alpha of zero are ignored, which reduces the number of compared, but also allows you to mask out which portions of the color set to use. The Selection image is collapsed to a 1D vector, as any spatial information in the Selection image is useless.

The Threshold specifies the maximum distance between the Input pixel and its closest color from the Selection image.  If the closest distance is greater than this value, the pixel returns black.  Remember, this is operating in float 3D space, so the maximum distance between 2 colors in a clamped 0-1 color cube is sqrt(3).  In unclamped colors, the distances could be much larger.

Rectify Output returns the absolute value of the distances.  Otherwise, the tool will output both positive and negative distances for each of the components of the vector.  Without this option, integer images would return black instead of the correct distances if negative.  We strongly recommend using KKey with float inputs.

The final option, Use cuda, is where the magic happens. If you don’t have CUDA, the tool will failover to a CPU only version, and will thus be slower, but that allows it to be more render-farm friendly.   This toggle is also good for benchmarking your CUDA hardware.

Download Comparing keyer geometry

Download KrazyKey keying example

So give the tool a try and leave some feedback in the comments section.  We’re also curious about the performance of various CUDA hardware, so let us know what kind of speed improvements you’re getting.