Performance settings are defined in the Performances section of the Predict Engine preferences. See the Engine Preferences section for more details.

Measured display fast loading

Denoiser level

Regularization

The correct amount of path samples, or equivalently the rendering time, required to obtain noise-free images depends on the scene content and, more specifically, on the complexity of the paths that light follows before reaching the camera.

For instance, indoor scenes take longer to render when most light comes from the external environment and must cross windows to lit the objects. In the same vein, caustics effects due to refractions or reflections of shiny surfaces might require more time to render smoothly as they involve two or more bounces between the camera and the light source.

Predict provides a feature called Regularization to help reducing the rendering time in such scenes, at the cost of extra bias in the resulting images. The introduced bias can be progressively removed to converge to exact results.

Optimized regularization settings:

• Bias level:

  • • Low: Low bias but noisy shiny paths,

    • Moderate: Low noise on shiny paths but additional bias,

    • Strong: Minimal noise on shiny paths but extra bias.

• Convergence speed: How fast the rendering converges to exact results:

  • • 0 (minimum allowed value): No convergence,

    • 0.5 (maximum allowed value): Convergence to exact results at maximal speed, at the cost of extra noise.

Fixed regularization settings:

• Attenuation factor: 

• • • 0 (minimum allowed value): Turn off regularization: The image is not biased but noise might be important on shiny paths

    • 1 (maximum allowed value): Noise is minimal but further bias is introduced

• Values between 0.005 and 0.1 are recommanded to accelerate the rendering of caustics while avoiding introducing too much bias (e.g. caustics on the bottom of a pool, light going through a glass of liquid, ...)

• Values above 0.1 are recommanded for scenes with indirect illumination (e.g. indoor scenes behind windows, ...)

• Convergence speed: How fast the rendering converges to exact results:

• • • 0 (minimum allowed value): No convergence

    • 0.5 (maximum allowed value): Convergence to exact results at maximal speed, at the cost of extra noise on shiny paths

Further details on regularization:

• Weier et al., Optimised Path Space Regularization, 2021

• Jendersie, J. and Grosch, T., Microfacet Model Regularization for Robust Light Transport. Computer Graphics Forum, 38: 39-47. https://doi.org/10.1111/cgf.13768

• Kaplanyan, A. and Dachsbacher, C., Path Space Regularization for Holistic and Robust Light Transport. Comput. Graph. Forum 32(2): 63-72 (2013) 

Russian Roulette