Page 17 of 25
Re: CGI tech news box
Posted: Thu Aug 29, 2019 6:40 am
by lacilaci
epilectrolytics wrote: ↑Thu Aug 29, 2019 6:10 am
We badly need a solution for SDS-paths in LuxCore as well as faster unbiased rendering for technical renders (laser, volumes etc).
I don't know how fast BiDirVM will be once it works but if path guiding can help with that it will be very welcome for sure.
Also I think there is much more potential in hybrid rendering, that is splitting the render tasks along "ray classes" or something and have different devices have a go at it with different algorithms.
I agree... But looking at the video now, it seems that Dade is already very close to what corona is doing (combinig light tracing and pathtracing)
Re: CGI tech news box
Posted: Thu Aug 29, 2019 8:21 am
by Mango3
lacilaci wrote: ↑Thu Aug 29, 2019 5:56 am
Well you're comparing caching and guiding as if they're competing for which does a better job.
Afaik these features aren't mutually exclusive, corona does pathguiding and uses caching. Also both vray's lightcache and corona's uhd cache don't suffer from lightleaks and can be used without problems for animations. They are both valuable, especially if they can be used both at the same time.
Biggest issue is pathguiding not being able to run on gpus. So unless pathguiding can bring same priced cpu ahead of same priced gpu in same generation in terms of performance, it just might be worth to explore other ways for getting photongi better and faster for example...
On another hand, luxcore already can already do hybrid rendering where cpu does lighttracing independently for caustics. Maybe having pathguiding for the cpu lighttracing part could give us caustics+sds on cpu while gpu does pathtracing?
Sadly of course in luxcore I've noticed heavy problems with leaking and odd reflections in more complex interiors with photongi.
Very thin objects (thin walls and floorboards on ground or wall decorations) can show a lot of problems no matter the settings.
And reflections of distant objects also tend to show weird variable bright spots.
Not really, I dint want to imply that they are competing or mutually exclusive.
There are several versions of path guiding. Of course one could combine path guiding for direct illumination with caching for secondary GI or have guided photon tracing/emission.
Other improvements deal with better light selection. Corona uses an adaptive method for direct illumination sampling:
https://cgg.mff.cuni.cz/~jaroslav/paper ... /index.htm This can be used on top of path guiding for instance.
The problems with light leaking and animation flickers can be worked around with retracing of GI near corners, like V-Ray does, and precomputing the GI cache for the whole camera path instead of just the current view. But ofc, Chaos Group has 350 employees and can afford to have engineers find solutions to work around these issues.
Re: CGI tech news box
Posted: Thu Aug 29, 2019 9:11 am
by Dade
Quite funny as it is exactly what I have done (with the additional scenario of heterogeneous CPU/GPU workload). The only missing part is the SDS paths and, while they seem to have used vertex merging, I'm planning, at the moment, to use path space regularization:
https://cg.ivd.kit.edu/publications/p20 ... n_2013.pdf
Re: CGI tech news box
Posted: Thu Aug 29, 2019 9:37 am
by lacilaci
Dade wrote: ↑Thu Aug 29, 2019 9:11 am
Quite funny as it is exactly what I have done (with the additional scenario of heterogeneous CPU/GPU workload). The only missing part is the SDS paths and, while they seem to have used vertex merging, I'm planning, at the moment, to use path space regularization:
https://cg.ivd.kit.edu/publications/p20 ... n_2013.pdf
Ondra mentioned vcm as not very good solution and also that they use machine learning at some point in the process... There should be some more technical talk somewhere about this...
Re: CGI tech news box
Posted: Thu Aug 29, 2019 10:01 am
by lacilaci
lacilaci wrote: ↑Thu Aug 29, 2019 9:37 am
Dade wrote: ↑Thu Aug 29, 2019 9:11 am
Quite funny as it is exactly what I have done (with the additional scenario of heterogeneous CPU/GPU workload). The only missing part is the SDS paths and, while they seem to have used vertex merging, I'm planning, at the moment, to use path space regularization:
https://cg.ivd.kit.edu/publications/p20 ... n_2013.pdf
Ondra mentioned vcm as not very good solution and also that they use machine learning at some point in the process... There should be some more technical talk somewhere about this...
Also note that even when they have caustics now, corona still does hybrid glass rendering by default and you need to switch to normal glass for caustics! It is simply not efficient to always render glass properly and very ugly always using archglass without refraction (which is giving me a lot of trouble in my new personal project where glass and water takes most of the scene)
Re: CGI tech news box
Posted: Thu Aug 29, 2019 10:22 am
by lacilaci
in this podcast Jaroslav Krivanek and Ondra Karlik talk about how their caustics are under the hood actually photon mapping + mlt with a lot of automatization
https://player.fm/series/cg-garage/epis ... a-caustics
Re: CGI tech news box
Posted: Thu Aug 29, 2019 3:31 pm
by Dade
Dade wrote: ↑Thu Aug 29, 2019 9:11 am
Quite funny as it is exactly what I have done (with the additional scenario of heterogeneous CPU/GPU workload). The only missing part is the SDS paths and, while they seem to have used vertex merging, I'm planning, at the moment, to use path space regularization:
https://cg.ivd.kit.edu/publications/p20 ... n_2013.pdf
Hacking together a "proof of concept" took me a lot of time
viewtopic.php?f=5&t=1369 
Re: CGI tech news box
Posted: Thu Aug 29, 2019 4:58 pm
by FarbigeWelt
Still, notation and math in papers like the one you linked is beyond my current math skills. However, the visible few simple graphics explained fast intention of their so called regularization.
When I scan read the paper I realized chemistry and path finding share commonalities. A chemical bond can be viewed as the largest probability of at least one of two atoms‘ electron‘s location, means space. Chemical bonding means, for some types of bonds, overlapping two electrons‘ most dense probability space capturing the electrons in an even denser probability space.
For path tracking this means to add an element to the paths increasing their intersection probability in certain circumstances when common intersection probability is close to zero.
To my understanding this can be done with small angle variances or thinking visual connection of narrow cones instead of lines.
Re: CGI tech news box
Posted: Mon Sep 09, 2019 9:42 am
by epilectrolytics
Here's something for better metropolis sampling on GPU:
https://pdfs.semanticscholar.org/2875/9 ... 983429.pdf
Edit: Probably irrelevant for LuxCore, quote:
"Another problem of our approach is that the contribution of a sample exponentially decreases with the depth. This leads to an optimal depth of 2 or 3 which in turn only generated 3 or 7 coherent samples."
Sorry for posting before reading completely
Re: CGI tech news box
Posted: Mon Sep 09, 2019 2:38 pm
by kintuX
Variance-Aware Multiple Importance Sampling - Jaroslav Křivánek
Many existing Monte Carlo methods rely on multiple importance sampling (MIS) to achieve robustness and versatility. Typically, the balance or power heuristics are used, mostly thanks to the seemingly strong guarantees on their variance. We show that these MIS heuristics are oblivious to the effect of certain variance reduction techniques like stratification. This shortcoming is particularly pronounced when unstratified and stratified techniques are combined (e.g., in a bidirectional path tracer). We propose to enhance the balance heuristic by injecting variance estimates of individual techniques, to reduce the variance of the combined estimator in such