Editing Photon mapping (section)

===Optimizations===
* To avoid emitting unneeded photons, the initial direction of the outgoing photons is often constrained. Instead of simply sending out photons in random directions, they are sent in the direction of a known object that is a desired photon manipulator to either focus or diffuse the light. There are many other refinements that can be made to the algorithm: for example, choosing the number of photons to send, and where and in what pattern to send them.  It would seem that emitting more photons in a specific direction would cause a higher density of photons to be stored in the photon map around the position where the photons hit, and thus measuring this density would give an inaccurate value for [[irradiance]].  This is true; however, the algorithm used to compute [[radiance]] does ''not'' depend on irradiance estimates.
* For soft indirect illumination, if the surface is [[Lambertian reflection|Lambertian]], then a technique known as [[irradiance caching]] may be used to interpolate values from previous calculations.
* To avoid unnecessary collision testing in direct illumination, shadow photons can be used.  During the photon mapping process, when a photon strikes a surface, in addition to the usual operations performed, a shadow photon is emitted in the same direction the original photon came from that goes all the way through the object.  The next object it collides with causes a shadow photon to be stored in the photon map.  Then during the direct illumination calculation, instead of sending out a ray from the surface to the light that tests collisions with objects, the photon map is queried for shadow photons.  If none are present, then the object has a clear line of sight to the light source and additional calculations can be avoided.
* To optimize image quality, particularly of caustics, Jensen recommends use of a cone filter.  Essentially, the filter gives weight to photons' contributions to radiance depending on how far they are from ray-surface intersections.  This can produce sharper images.
* [http://research.nvidia.com/publication/hardware-accelerated-global-illumination-image-space-photon-mapping Image space photon mapping] achieves real-time performance by computing the first and last scattering using a GPU rasterizer.