The first attempt made from two old scenes seemed to me as if the caustics were too weak. Then I realized that when rendering with indirect PGI indirect caustic paths don't connect to lights but end with a cache lookup.
Also rendering a reference with Metropolis took too long on a scene that complex.
So I made a test scene with a glass cube in a room with an opening in a wall to another room illuminated by an area light. Test results:
Neither light tracing nor caustic cache render indirect caustics
![Crying or Very Sad :cry:](./images/smilies/icon_cry.gif)
But Metropolis sampling does, better with PathCPU than BiDir
![Smile :)](./images/smilies/icon_e_smile.gif)
PathOCL Metro shows the best caustics, but more fireflies and doesn't really converge even when approaching 100.000 samples.
Also it renders way brighter with brightness change during render
![Question :?:](./images/smilies/icon_question.gif)
This makes me wish path guiding would be considered as the next render development project after Path Regularization is finished.
Because right now we are piling up quick fixes that cover single problems leaving others unsolved.
Like: PathOCL + PGI: Blazingly fast but unbiased and no good caustis
First fix: Hybrid Path with fast caustics but missing SDS
Second fix: Path Regularization for SDS (fast or slow, easy or tricky - yet unknown)
Still no indirect caustics.
Compared with this path guiding (Corona style) would be slower (CPU only) but faster than BiDir is now and cover all path type combinations.
Also it would work "under the hood" without requiring the user to decide which engine/sampler combination to use or long trial & error parameter tweaking sessions.