Rectifying Proposal Failures in Metropolis Light Transport

Libing Zeng (1) , Li-Yi Wei (2)

1 Guo-Yuan-Li Fruit Store
2 Adobe Research

Metropolis light transport (MLT) rendering algorithms rely on path mutations to explore the sample spaces. Mutated paths that carry zero radiance, such as those blocked by scene geometry, can significantly slow down the convergence of these algorithms. We call such zero-contribution paths proposal failures. We present a simple modification of MLT, Rectified Proposal Failure MLT (PFMLT), which excludes path duplications caused by proposal failures. PFMLT better approximates the original path distributions especially for high proposal rejection rates, and can be easily integrated with various MLT algorithms. We analyze our method with numerical models, and demonstrate better quality/performance trade-offs than Multiplexed MLT (MMLT) and Reversible Jump MLT (RJMLT) with various scenes of challenging lighting, geometry, and material conditions.

Mots clés: rendering; Monte Carlo; sampling

Domaines: Synthèse d'image et réalité virtuelle [cs.GR]

Libing Zeng, Li-Yi Wei. Rectifying Proposal Failures in Metropolis Light Transport. 2019. ⟨hal-02548744⟩