what the hell?! inverted refraction

Hi Guys!
Hope that’s the good part of the forum for this.
Here is a recent rendering ( cycle / principled bdsf) . I’ve got 2 strange problems.
->If you look for the refraction on the plate ( candle and glass), they are inverted, like projected on the plate but not in the good direction), same on the glasses.
Do you have an idea of the nature of the problem?!
-> The glasses are not transparent. if you put something behing the png, the glass are totally full… How strange too is that?!

YOu can see better here, whitout the glares…

  1. please, show the nodes & possibly the geometry/normals - tho all times best is to provide .blend (exemplary scene file)
  2. that’s by design, Transparent Glass option was only recently introduced, you can use daily builds if you wish so

that’s a pretty simple setup. For the glass it’s a classic glass shader+ an env map.

Your plate is concave and therefore this inverted reflection is quite expected… You’ll also see it in real world if you have a plate with that same concavity, and the objects are closer to the plate than its focal point.

Does principled do glass correctly though? If I have a bent plane such that I can see both sides at the same time, using a dielectric glossy principled setup, all looks nice. Meaning that IOR is inverted automatically for backfacing faces. That’s handy when doing regular stuff where you need to simplify using single faced geometry, but wrong when doing glass which needs entry and exit point. Increasing transmission does not seem to invert the IOR blending at a specific point.

What we need is a switch that lets us toggle single sided use for this material. If unchecked, backfacing normals should get inverted IOR which would make glass work fine.

Also, is the clearcoat roughness broken? Going from 0.0177 to 0.0178 causes an abrupt change in clearcoat (set to 1) output. Do I have a broken version or is this some weird design choice I don’t understand? I currently do all my clearcoats outside of principled, roughness is typically so low anyway I don’t need to do any complicated PBR stuff to compensate fresnell for roughness.