What i do...

I originally downloaded .gif and .jpg files from the internet as a substrate to begin manipulation on. I would crop them down to a good starting shape that would look interesting when mirrored back on itself at each edge. These days, i have a huge backlog of files i've created myself and these are mostly what i use to start a new image from.

Once a starting image is decided on, i use blurring routines to spread data from adjacent points into each other. This is followed by a convolution of the color intensity function used to display the image. This involves making dimmer pixels brighter and some of the middling brighter ones dimmer -- or some similar transformation. Each convolution is fed back into the blurring routines in a continuous cycle. It is possible to carefully control the rate of spreading, so that structures will spontaneously arise and propagate through the resulting image display.

Colors can be mapped into different values as desired; they can be added or subtracted in intensity and saturation as well as being rotated around the color wheel to differing hues. The combinations of red green and blue that are added to make each color are independantly controlable to result (with some considerable amounts of practice) in absolute control of color placement over the spatial lattice of the image.

The overall effect is like creating a fourier transform on the original image that warps it kaleidoscopically across color and image space into a totally new, (frequently) unpredictable, and (with practice) eye-pleasing configuration. With appropriate cropping, subsets of the image can be selected for further manipulation, frequently resulting in shapes that only have the barest resemblance to the original image.