Hybrid Bidir

[A-man]

Is the high RAM usage the video card RAM or the CPU RAM? I'm curious because my card only has 1GB RAM

[Lord Crc]

GPU ram is used to store the geometry, without instances it goes quickly.

[SATtva]

Is the high RAM usage the video card RAM or the CPU RAM? I'm curious because my card only has 1GB RAM

What J the Ninja been talking here is system RAM which is used to keep the state of the integrator.

[J the Ninja]

Here's the latest comparison with the "locked tower" scene:

Sampler:

Hybrid:

Hybrid bidir still seems to be leaving some odd dark areas on the glasses, mirror, and plant leaves?

[Dade]

Hybrid bidir still seems to be leaving some odd dark areas on the glasses, mirror, and plant leaves?

Very likely, at the moment I have done tests only with a LuxBall5 scene made only of matte materials (and now the rendering should be correct in that case). I'm also aware that it is unable to correctly render Psor's cube.

I have lately fixed the path weights (in order to produce correct renderings), a couple of other bugs and written the code to connect the light path directly with the camera (i.e. a basic requirement to render caustics).

Side note: BiDir can be seen as blend of different rendering methods: path tracing, light tracing, etc. By using constant path weights I'm effectively taking the worst (i.e. noise) out of all rendering methods
MiS solves this problem (by picking the best instead) but it is the very last step I'm planning (i.e. don't worry too much about the noise in current hybrid BiDir renderings). Before MiS, I want to get everything else right.

[J the Ninja]

One step at a time.

Thanks for the update and your work, Dade

[B.Y.O.B.]

One step at a time.

+1
I wanted to thank you, too.

[jeanphi]

MiS solves this problem (by picking the best instead) but it is the very last step I'm planning (i.e. don't worry too much about the noise in current hybrid BiDir renderings). Before MiS, I want to get everything else right.

Wise choice. It took me months if not years to get it right, and it requires storing much more data to compute the weighting factors. Hopefully we now have a reference implementation which I was lacking at that time so it should be easier, but it remains a tough algorithm. Best of luck! And feel free to ask me if you need help in there. I've not had the time to deeply look into your latest work but it seems to be shaping up nicely.

Jeanphi

[Dade]

Be careful when doing tests with hybrid BiDir: I have spent some time trying to understand why Psor's cube wasn't correctly rendered. At the end, I just realized the scene was made mostly by quads (i.e. not triangles). This was leading to some very wired result because quads were intersected when tracing rays on the CPU (i.e. eye and light path) while they weren't when tracing connection rays on the GPU .

I will add some kind of check/warning because it can be very confusing.