Dif­fer­en­ti­able path tra­cing of­fers a prin­cipled route to re­cov­er­ing phys­ic­al ma­ter­i­al and light­ing para­met­ers, but the com­bin­a­tion of high vari­ance and poor nu­mer­ic­al con­di­tion­ing of­ten makes it too brittle to use in prac­tice. This is es­pe­cially the case when light­ing is al­to­geth­er un­known, or when the scene con­tains com­plex light trans­port ef­fects. Pri­or work re­cently showed that the vari­ance re­duc­tion provided by a ra­di­ance cache can al­le­vi­ate these chal­lenges.

We re­vis­it the com­bin­a­tion of in­verse ren­der­ing and ra­di­ance cach­ing with a twist, by in­tro­du­cing a spa­tial blend­ing field that loc­ally in­ter­pol­ates between the cache and stand­ard un­biased es­tim­at­ors. Re­curs­ive ap­plic­a­tion of this idea yields a rich design space of eval­u­ation strategies and inter-es­tim­at­or con­sist­ency losses; we map this space and identi­fy ef­fect­ive com­pon­ents. A sur­pris­ing prop­erty of the res­ult­ing al­gorithm is that it can ac­cur­ately re­cov­er ma­ter­i­al para­met­ers even when the light­ing is not uniquely iden­ti­fi­able from the ob­ser­va­tions. Our ex­per­i­ments demon­strate sig­ni­fic­ant im­prove­ments in speed and ro­bust­ness over pri­or work, mak­ing a strong case for in­clud­ing ra­di­ance cach­ing as a stand­ard com­pon­ent of fu­ture phys­ic­ally based in­verse ren­der­ing sys­tems.