Understanding how primary and secondary samples are related helps estimate the maximum number of rays Karma uses when rendering. User settings, per-object overrides, and algorithms make predicting the exact number impossible. Understanding how these parameters influence one another helps debug performance issues.

The number of rays Karma uses is determined by the total number of primary samples and secondary samples fired for each pixel. The selected convergence mode also affects this estimate, however pixel oracles only affect the number of primary samples.

[Primary Samples] + [Secondary Samples]

[Primary Samples] *
([Max Direct Samples] + [Max Indirect Samples])

[Number of Lights Sampled] *
[Light Sample Quality] * [Max Secondary Samples]

[Max Secondary Samples] * [Diffuse Quality] +
[Max Secondary Samples] * [Reflection Quality] +
[Max Secondary Samples] * [Refraction Quality] +
[Max Secondary Samples] * [Volume Quality] +
[Max Secondary Samples] * [SSS Quality]

[Primary Samples] + [Secondary Samples]

[Primary Samples] *
([Max Direct Samples] + [Max Indirect Samples])

[Number of Lights Sampled] *
[Light Sample Quality]

[1 Randomly-Chosen Indirect Sample]

Oversampling (sending out more rays than you need for a given quality level) and under-sampling (sending out too few rays to achieve a given quality level) are sources of slow-downs when rendering. Understanding the differences between primary and secondary samples will help you to use Karma more effectively.

Karma supports adaptive primary samples, using pixel oracles. Pixel oracles represent the logic Karma should use when casting rays from each pixel. Currently, Karma ships with 2 pixel oracles: Variance and Uniform.

By default, Karma uses the Variance oracle, which detects differences between neighboring pixels. This is to avoid sending too many primary samples by stopping when each additional sample will have very little effect on the final pixel color. Karma will send a small number of primary samples, then the Variance Oracle starts comparing pixels against the Oracle’s Variance Threshold. Once the variance between the previous samples of the current pixel and adjacent pixels fall below that threshold, Karma stops firing primary samples for that pixel.

The Uniform Oracle simply sends the same number of primary samples from every pixel in the scene. You may want to switch to the Uniform Oracle when extreme depth of field (DOF) or motion blur dominates the image. In those cases, Karma ends up fully sampling every pixel anyway, so measuring variance just adds overhead.

The convergence mode is the strategy used by Karma to decide when to send secondary samples into the scene. Karma CPU supports two convergence modes: Automatic and Path Traced, while Karma XPU only supports Path Traced mode. While convergence mode specifically affects secondary rays, it can affect how many primary samples are needed, especially for Karma CPU.

The next chapters have more information on convergence modes for Karma CPU and Karma XPU specifically.

To avoid rays potentially traveling forever, Karma provides per-type “bounce” limits. You may need to increasing these limits to ensure Karma takes enough indirect samples to get the desired look. Refraction often needs a higher limit, to improve the quality of scenes with glass and liquid materials. Volumes may also benefit from increased limits, to capture the subtle effects of volumetric lighting.

Increasing the limits can increase render times substantially, and require higher numbers of rays to resolve noise. Be careful to only increase limits when needed to achieve a desired look.

Limits Render Properties

Diffuse Limit

The number of times diffuse rays can propagate through your scene.

Unlike the Reflect and Refract Limits, this parameter will increase the overall amount of light in your scene and contribute to the majority of global illumination. With this parameter set above zero diffuse surfaces will accumulate light from other objects in addition to direct light sources.

In this example, increasing the Diffuse Limit has a dramatic effect on the appearance of the final image. To replicate realistic lighting conditions, it is often necessary to increase the Diffuse Limit. However, since the amount of light contribution usually decreases with each diffuse bounce, increasing the Diffuse Limit beyond 4 does little to improve the visual fidelity of a scene. Additionally, increasing the Diffuse Limit can dramatically increase noise levels and render times.

This is a float because all limits are stochastically picked per-sample, so for example you can set the diffuse limit to 3.25 and have 25% of the rays with a diffuse limit of 4 and 75% of rays with a diffuse limit of 3.

Reflection Limit

The number of times a ray can be reflected in your scene.

This example shows a classic “Hall of Mirrors” scenario with the subject placed between two mirrors.

This effectively creates an infinite series of reflections.

Remember that the first time a light source is reflected in an object, it is considered a direct reflection. Therefore, even with Reflect Limit set to 0, you will still see specular reflections of light sources.

From this camera angle the reflection limits are very obvious and have a large impact on the accuracy of the final image. However, in most cases the reflection limit will be more subtle, allowing you to reduce the number of reflections in your scene and optimize the time it takes to render them.

This is a float because all limits are stochastically picked per-sample, so for example you can set the diffuse limit to 3.25 and have 25% of the rays with a diffuse limit of 4 and 75% of rays with a diffuse limit of 3.

Refraction Limit

This parameter control the number of times a ray be refracted in your scene.

This example shows a simple scene with ten grids all in a row.

By applying a refractive shader, we will be able see through the grids to an image of a sunset in the background.

From this camera angle, in order for the image to be accurate, the refraction limit must match the number of grids that that are in the scene. However, most scenes will not have this number of refractive objects all in a row and so it is possible to reduce the refract limit without affecting the final image while also reducing the time it takes to render them.

Keep in mind that this Refract Limit refers to the number of surfaces that the ray must travel through, not the number of objects.

Remember that the first time a light source is refracted through a surface, it is considered a direct refraction. Therefore, even with Refract Limit set to 0, you will see refractions of Light Sources. However, since most objects in your scene will have at least two surfaces between it and the light source, direct refractions are often not evident in your final render.

This is a float because all limits are stochastically picked per-sample, so for example you can set the diffuse limit to 3.25 and have 25% of the rays with a diffuse limit of 4 and 75% of rays with a diffuse limit of 3.

Volume Limit

The number of times a volumetric ray can propagate through a scene. It functions in a similar fashion to the Diffuse Limit parameter.

Increasing the Volume Limit parameter will result in much more realistic volumetric effects. This is especially noticeable in situations where only part of a volume is receiving direct lighting. Also, in order for a volumetric object to receive indirect light from other objects, the Volume Limit parameter must be set above 0.

With the Volume Limit set to values above zero, the fog volume takes on the characteristic light scattering you would expect from light traveling through a volume. However, as with the Diffuse Limit, the light contribution generally decreases with each bounced ray and therefore using values above 4 does not necessarily result in a noticeably more realistic image.

Also, increasing the value of this parameter can dramatically increase the amount of time spent rendering volumetric images.

This is a float because all limits are stochastically picked per-sample, so for example you can set the diffuse limit to 3.25 and have 25% of the rays with a diffuse limit of 4 and 75% of rays with a diffuse limit of 3.

SSS Limit

The number of times a SSS ray can propagate through a scene. It functions in a similar fashion to the Diffuse Limit parameter.

This is a float because all limits are stochastically picked per-sample, so for example you can set the diffuse limit to 3.25 and have 25% of the rays with a diffuse limit of 4 and 75% of rays with a diffuse limit of 3.

Many of the sampling settings for each render delegate can also be set on objects of the scene. Per-light and per-object settings are usually multpliers on their global counterpart. If a property does not mention “scale” or “multiplies” in its description, then it overrides the equivalent global property.

Also it’s worth noting that if a render delegate or convergence mode does not support a given sampling style/feature, per object overrides won’t have an effect.

To understand sampling and noise better, follow the next chapter for your render delegate of choise.