3.3.4 Radiosity

3.3.4.1 How Radiosity Works

The problem of ray-tracing is to figure out what the light level is at each point that you can see in a scene. Traditionally, in ray tracing, this is broken into the sum of these components:

Diffuse
the effect that makes the side of things facing the light brighter;
Specular
the effect that makes shiny things have dings or sparkles on them;
Reflection
the effect that mirrors give; and
Ambient
the general all-over light level that any scene has, which keeps things in shadow from being pure black.

POV-Ray's radiosity system, based on a method by Greg Ward, provides a way to replace the last term - the constant ambient light value - with a light level which is based on what surfaces are nearby and how bright in turn they are.

The first thing you might notice about this definition is that it is circular: the brightness and color of everything is dependent on everything else and vice versa. This is true in real life but in the world of ray-tracing, we can make an approximation. The approximation that is used is: the objects you are looking at have their ambient values calculated for you by checking the other objects nearby. When those objects are checked during this process, however, their diffuse term is used. The brightness of radiosity in POV-Ray is based on two things:

  1. the amount of light "gathered"
  2. the 'diffuse' property of the surface finish

An object can have both radiosity and an ambient term. However, it is suggested that if you use radiosity in a scene, you either set ambient_light to 0 in global_settings, or use ambient 0 in each object's finish. This lighting model is much more realistic, and POV-Ray will not try to adjust the overall brightness of the radiosity to match the ambient level specified by the user.

How does POV-Ray calculate the ambient term for each point? By sending out more rays, in many different directions, and averaging the results. A typical point might use 200 or more rays to calculate its ambient light level correctly.

Now this sounds like it would make the ray-tracer 200 times slower. This is true, except that the software takes advantage of the fact that ambient light levels change quite slowly (remember, shadows are calculated separately, so sharp shadow edges are not a problem). Therefore, these extra rays are sent out only once in a while (about 1 time in 50), then these calculated values are saved and reused for nearby pixels in the image when possible.

This process of saving and reusing values is what causes the need for a variety of tuning parameters, so you can get the scene to look just the way you want.

3.3.4.2 Adjusting Radiosity

As described earlier, radiosity is turned on by using the radiosity{} block in global_setting. Radiosity has many parameters that are specified as follows:

    global_settings { radiosity { [RADIOSITY_ITEMS...] } }
RADIOSITY_ITEMS:
    adc_bailout Float | always_sample Bool | brightness Float | 
    count Integer | error_bound Float | gray_threshold Float |
    load_file Filename | low_error_factor Float | max_sample Float |
    media Bool | minimum_reuse Float | nearest_count Integer |
    normal Bool | pretrace_end Float | pretrace_start Float |
    recursion_limit Integer | save_file Filename

Each item is optional and may appear in any order. If an item is specified more than once the last setting overrides previous values. Details on each item is given in the following sections.

Note: Considerable changes have been made to the way radiosity works in POV-Ray 3.5 compared to POV-Ray 3.1. Old scene will not render to the same result, if they render at all. It is not possible to use the #version directive to get backward compatibility for radiosity.

3.3.4.2.1 radiosity adc_bailout

You can specify an adc_bailout for radiosity rays. Use adc_bailout = 0.01 / brightest_ambient_object for good results. Default is 0.01.

3.3.4.2.2 always_sample

You can force POV-Ray to only use the data from the pretrace step and not gather any new samples during the final radiosity pass. This may reduce splotchiness. To do this, use always_sample off, by default it is on. It can also be usefully when reusing previously saved radiosity data.

3.3.4.2.3 brightness

The brightness keyword specifies a float value that is the degree to which objects are brightened before being returned upwards to the rest of the system. The default value is 1.0. In cases where you would raise the global_settings{ambient_light value} to increase the over all brightness in a non-radiosity scene, you can use brightness in a radiosity scene.

3.3.4.2.4 count

The integer number of rays that are sent out whenever a new radiosity value has to be calculated is given by count. A value of 35 is the default, the maximum is 1600. When this value is too low, the light level will tend to look a little bit blotchy, as if the surfaces you are looking at were slightly warped. If this is not important to your scene (as in the case that you have a bump map or if you have a strong texture) then by all means use a lower number.

3.3.4.2.5 error_bound

The error_bound float value is one of the two main speed/quality tuning values (the other is of course the number of rays shot). In an ideal world, this would be the only value needed. It is intended to mean the fraction of error tolerated. For example, if it were set to 1 the algorithm would not calculate a new value until the error on the last one was estimated at as high as 100%. Ignoring the error introduced by rotation for the moment, on flat surfaces this is equal to the fraction of the reuse distance, which in turn is the distance to the closest item hit. If you have an old sample on the floor 10 inches from a wall, an error bound of 0.5 will get you a new sample at a distance of about 5 inches from the wall.

The default value of 1.8 is good for a smooth general lighting effect. Using lower values is more accurate, but it will strongly increase the danger of artifacts and therefore require higher count. You can use values even lower than 0.1 but both render time and memory use can become extremely high then.

3.3.4.2.6 gray_threshold

Diffusely interreflected light is a function of the objects around the point in question. Since this is recursively defined to millions of levels of recursion, in any real life scene, every point is illuminated at least in part by every other part of the scene. Since we cannot afford to compute this, if we only do one bounce, the calculated ambient light is very strongly affected by the colors of the objects near it. This is known as color bleed and it really happens but not as much as this calculation method would have you believe. The gray_threshold float value grays it down a little, to make your scene more believable. A value of .6 means to calculate the ambient value as 60% of the equivalent gray value calculated, plus 40% of the actual value calculated. At 0%, this feature does nothing. At 100%, you always get white/gray ambient light, with no hue.

Note: this does not change the lightness/darkness, only the strength of hue/grayness (in HLS terms, it changes S only). The default value is 0.0

3.3.4.2.7 low_error_factor

If you calculate just enough samples, but no more, you will get an image which has slightly blotchy lighting. What you want is just a few extra interspersed, so that the blending will be nice and smooth. The solution to this is the mosaic preview, controlled by pretrace: it goes over the image one or more times beforehand, calculating radiosity values. To ensure that you get a few extra, the radiosity algorithm lowers the error bound during the pre-final passes, then sets it back just before the final pass. The low_error_factor is a float tuning value which sets the amount that the error bound is dropped during the preliminary image passes. If your low error factor is 0.8 and your error bound is set to 0.4 it will really use an error bound of 0.32 during the first passes and 0.4 on the final pass. The default value is 0.5.

3.3.4.2.8 max_sample

Sometimes there can be problems with splotchiness that is caused by objects that are very bright. This can be sometimes avoided by using the max_sample keyword. max_sample takes a float parameter which specifies the brightest that any gathered sample is allowed to be. Any samples brighter than this will have their brightness decreased (without affecting color). Specifying a non-positive value for max_sample will allow any brightness of samples (which is the default).

3.3.4.2.9 Media and Radiosity

Radiosity estimation can be affected by media. To enable this feature, add media on to the radiosity{} block. The default is off

3.3.4.2.10 minimum_reuse

The minimum effective radius ratio is set by minimum_reuse float value. This is the fraction of the screen width which sets the minimum radius of reuse for each sample point (actually, it is the fraction of the distance from the eye but the two are roughly equal). For example, if the value is 0.02, the radius of maximum reuse for every sample is set to whatever ground distance corresponds to 2% of the width of the screen. Imagine you sent a ray off to the horizon and it hits the ground at a distance of 100 miles from your eye point. The reuse distance for that sample will be set to 2 miles. At a resolution of 300*400 this will correspond to (very roughly) 8 pixels. The theory is that you do not want to calculate values for every pixel into every crevice everywhere in the scene, it will take too long. This sets a minimum bound for the reuse. If this value is too low, (which it should be in theory) rendering gets slow, and inside corners can get a little grainy. If it is set too high, you do not get the natural darkening of illumination near inside edges, since it reuses. At values higher than 2% you start getting more just plain errors, like reusing the illumination of the open table underneath the apple. Remember that this is a unit less ratio. The default value is 0.015.

3.3.4.2.11 nearest_count

The nearest_count integer value is the minimum number of old ambient values blended together to create a new interpolated value. The total number blended will vary depending on error_bound. All previous values that fit within the specified error_bound will be used in the average.

It will always be the n geometrically closest reusable points that get used. If you go lower than 4, things can get pretty patchy. This can be good for debugging, though. Must be no more than 20, since that is the size of the array allocated. The default value is 5.

3.3.4.2.12 Normal and Radiosity

Radiosity estimation can be affected by normals. To enable this feature, add normal on to the radiosity{} block. The default is off

3.3.4.2.13 Pretrace

To control the radiosity pre-trace gathering step, use the keywords pretrace_start and pretrace_end within the radiosity{} block. Each of these is followed by a decimal value between 0.0 and 1.0 which specifies the size of the blocks in the mosaic preview as a percentage of the image size. The defaults are 0.08 for pretrace_start and 0.04 for pretrace_end

3.3.4.2.14 recursion_limit

The recursion_limit is an integer value which determines how many recursion levels are used to calculate the diffuse inter-reflection. The default value is 3, the upper limit is 20.

3.3.4.2.15 Save and load radiosity data

You can save the radiosity data using save_file "file_name" and load the same data later using load_file "file_name". In general, it is not a good idea to save and load radiosity data if scene objects are moving. Even if data are loaded, more samples may be taken during rendering (which produces a better approximation). You can disable samples from being taken during the final rendering phase by specifying always_sample off.

3.3.4.3 Tips on Radiosity

Have a look at the "Radiosity Tutorial" in the "Advanced Tutorial" section, to get a feel for what the visual result of changing radiosity parameters is.

If you want to see where your values are being calculated set radiosity count down to about 20, set radiosity nearest_count to 1 and set gray_threshold to 0. This will make everything maximally patchy, so you will be able to see the borders between patches. There will have been a radiosity calculation at the center of most patches. As a bonus, this is quick to run. You can then change the error_bound up and down to see how it changes things. Likewise modify minimum_reuse.

One way to get extra smooth results: crank up the sample count (we have gone as high as 1300) and drop the low_error_factor to something small like 0.6. Bump up the nearest_count to 7 or 8. This will get better values, and more of them, then interpolate among more of them on the last pass. This is not for people with a lack of patience since it is like a squared function. If your blotchiness is only in certain corners or near certain objects try tuning the error bound instead. Never drop it by more than a little at a time, since the run time will get very long.

Sometimes extra samples are taken during the final rendering pass. These newer samples can cause discontinuities in the radiosity in some scenes. To decrease these artefacts, use a pretrace_end of 0.04 (or even 0.02 if you are really patient and picky). This will cause the majority of the samples to be taken during the preview passes, and decrease the artefacts created during the final rendering pass. You can force POV-Ray to only use the data from the pretrace step and not gather any new samples during the final radiosity pass. To do this, use "always_sample no" within the radiosity block inside global_settings.

If your scene uses ambient objects (especially small ambient objects) as light sources, you should probably use a higher count (100-150 and higher). For such scenes, an error_bound of 1.0 is usually good. Higher causes too much error, but lower causes very slow rendering. And it is important to adapt adc_bailout.