I have found this paper that maybe could give us ideas about this important issue in the YafaRay engine. YafaRay implementation of photon mapping techniques in its two versions diffuse and caustics, lacks an important element that was always considered in Jensen's original paper about photonmapping, which is the fact that light sources should know areas of importance to shoot photons at. The way YafaRay works now, photons can bounce in areas with little relevance to the final result, and photonmapping like other GI algorithms is foremost a problem of density. With lack of density in the photonmap we have low frequency noise and shadow flickering. Not only the traditional photonmapping integrator has got this problem, also the issues we have with SPPM are pretty much derived from this. Jensen proposed in his papers creating projection maps, which are typically implemented as bitmaps warped onto a bounding shape for the light source, where each bit determines whether there is geometry of importance in that direction.
Anyway I have found this paper that proposes a preliminar pass that consist in a construction of an importance map by using a rough path tracing estimation with origin at the light sources. The paper link is below: