While playing around in Blender, I discovered an interesting effect. This may not be news to everyone, but it was news to me.
If you give a light source a negative color value (where the red, green and blue values are less than 0), you get a kind of negative/inverse light source. For example, if you give an object with an emission node a color value of [0, -1, 0], you will get an inverse light that absorbs green light and leaves red and blue light.
Of course, this has no effect without other regular light sources, either lamps or light emitting objects. So, the interesting stuff happens when you place a regular light and an "inverse" light in a scene.
Now, you cannot simply assign a negative value to anything that accepts a color. You cannot even directly assign a negative value to a Combine RGB node.
There a number of ways you can get a negative color value. However, the most straight-forward method (I think) is this:
The only thing you have to keep in mind is that you are no longer selecting the color you want to see. Instead, you are selecting how much of the Red, Green, and Blue light in the scene you want to cancel out. Otherwise, the light acts in much the same way. The strength will work as expected.
In Figure 2, there are only three objects, two spheres and a plane. The plane has a pure white Diffuse node. The left sphere has a pure white Emission node, strength 3. The right sphere has the same Emission node, but the Mix RGB node (as shown in Figure 1) is plugged into the Emission color socket. The world color is set to pure black, so you can see that the right sphere is not emitting any regular light, but it is emitting a negative light that cancels out all of the green light and some of the blue light it encounters.
If I bump up the strength of the right sphere's Emission node to 30, you get this:
There is a side effect of this process: you get a lot of noise. Figure 2 and Figure 3 were set to 100 samples, and clamping set to 5. Here is one with 1000 samples:
The noise is reduced, but not eliminated. The same scene without the right sphere is pretty smooth at 1000 samples. I'm sure there is some math that would explain why the inverse light causes extra noise, but I haven't worked it out.
However, switching to Branched Path Tracing in the Sampling panel will allow you to smooth things out, without having a crazy amount of samples - but, at the cost of rendering speed, I think. Using these settings for the Branched Path Tracing options:
and using the first example, with the inverse light set to an Emission strength of 3, you get this:
Still noisy, but getting better.
I chose, for now, to just deal with the noise. Some images I may smooth out in Photoshop, while I may leave the noise in other images.
Here are some examples of abstracts I created using this process: