Pixels on the screen can have colors between 0 and 1. It’s common when rendering particles with additive blending that the result of the addition is over 1.0, in which case the value is simply clamped to 1. This looks really bad.
Here’s a picture of a glowy circle being rendered on top of a gray background. The background has the color (0.5, 0.5, 0.5) while the circle has the color (1, 1, 1) at the center and fades out to (0, 0, 0) at the edges. The circle is rendered with additive blending ([icode]glBlendFunc(GL_ONE, GL_ONE)[/icode] = simply add the color on the screen together with the color we’re rendering).
You’d have to be blind to not notice the big white circle in the middle where the color was clamped to 1.0. In fact, the center is actually (1.5, 1.5, 1.5) after the additive blending!
Many 3D games use high dynamic range rendering to solve this. They essentially render to a render target that can hold values over 1.0, and then run the final color through a tone mapping function to convert it to a value between 0.0 and 1.0 again. The main goal of this is to convey the difference in brightness to the user better. Currently, the player sees the whole clamped area as 1.0 when the middle part is actually 50% brighter than the edge of the clamped area. The simplest tone mapping function is f(x) = x / (x+1), which has this graph:
As you can see, the value goes towards 1.0, but it never actually reaches it. That means that regardless of how bright a pixel is, there’s always a brighter color available (theoretically at least). That allows us to represent values that are brighter than 1.0, allowing us to differ between a somewhat bright pixel with an intensity of 1.0 and looking into the sun which has an intensity in the hundreds.
Unless we start meddling with frame buffer objects, HDR textures and tone mapping shaders, we can’t do this the right way. What we can do is mess with the blending function in an attempt to get a behavior similar to the above function. Here’s the exact same scene as above, but this time I’m using [icode]glBlendFunc(GL_ONE_MINUS_DST_COLOR, GL_ONE)[/icode] instead.
This blend mode works a bit differently. Instead of simply adding together the incoming colors, we don’t let them go over 1.0. This is achieved by scaling the circle’s color with how much we have left until we reach 1.0 which is (1 - destination color). Here’s an image displaying the process:
With this blend mode, if we were to repeatedly blend in 0.5 onto the screen we’d end up with the following colors after X iterations:
0: 0.0
1: 0.5
2: 0.75
3: 0.875
4: 0.9375
Our screen can no longer be saturated to 1.0 unless we explicitly write 1.0 to the screen!
Here’s a comparison between [icode]glBlendFunc(GL_ONE, GL_ONE)[/icode] and [icode]glBlendFunc(GL_ONE_MINUS_DST_COLOR, GL_ONE)[/icode] when rendering 5 of the above circles on top of each other.
[icode]glBlendFunc(GL_ONE, GL_ONE)[/icode]:
[icode]glBlendFunc(GL_ONE_MINUS_DST_COLOR, GL_ONE)[/icode]:
Although the normal additive blending actually is brighter on the screen since most of the pixels are (1, 1, 1), using GL_ONE_MINUS_DST_COLOR better conveys the fact that those superbright pixels have very different intensities.
EDIT:
Here’s a more interesting example where the center of the circle has the color (1.0, 0.5, 0.25) and 10 circles stacked together:
[icode]glBlendFunc(GL_ONE, GL_ONE)[/icode]:
[icode]glBlendFunc(GL_ONE_MINUS_DST_COLOR, GL_ONE)[/icode]:
