You know, I used to get angry also like you trying to demonstrate people how something works, and as I have learnt through the years of my life, is let people live in their own stupidity.
Remeber your position and where others are.
Shut these guys up!
-
joegiampaoli
- Posts: 837
- Joined: Thu Oct 05, 2006 7:12 am
- Location: San Miguel de Allende-MEXICO
- Contact:
-
joegiampaoli
- Posts: 837
- Joined: Thu Oct 05, 2006 7:12 am
- Location: San Miguel de Allende-MEXICO
- Contact:
Users have latched onto the term "MLT" as representing unbiased ray tracing, and that any other method is biased. This is not true. There are several variants of unbiased ray (or path) tracing. MLT is one of them. It is neither higher tech nor lower tech than other variants. Unbiased is unbiased and each variant has its strengths and weaknesses.
It is a common misconception that all unbiased engines are using MLT. I don't believe anyone is claiming to use MLT as some sort of "brand" advantage.
It is a common misconception that all unbiased engines are using MLT. I don't believe anyone is claiming to use MLT as some sort of "brand" advantage.
hmm, i don't know if that was directed at me, but i didn't claim only mlt is unbiased. unbiased-ness isn't a really obvious thing to understand, and anyone looking to clarify the point should read keenan crane's concise writeup here: http://www.cgafaq.info/wiki/Bias_in_rendering
i do, however, feel that mlt is "higher tech" than other methods. it is a very interesting way to perform integration, in that it allows the limited use of context (through correlated samples), and a very powerful form of importance sampling in difficult lighting situations such as the reflected caustics i mentioned. the mutation scheme described by kelemen et al. is particularly powerful because it doesn't obsessively focus on bright regions, through the use of multiple importance sampling between mlt and "pure" monte carlo methods (a stroke of genius). of course i'm biased *drumroll* because i've totally fallen in love with the elegance and power of the method, but it can't be denied that it confers significant benefits in difficult sampling situations.
so, whether or not mlt is being used as a marketing term, the fact is that it forms a solid basis for the current generation of rendering engines, and it's not unreasonable for users to associate "implements mlt" with "efficiently renders almost anything".
i do, however, feel that mlt is "higher tech" than other methods. it is a very interesting way to perform integration, in that it allows the limited use of context (through correlated samples), and a very powerful form of importance sampling in difficult lighting situations such as the reflected caustics i mentioned. the mutation scheme described by kelemen et al. is particularly powerful because it doesn't obsessively focus on bright regions, through the use of multiple importance sampling between mlt and "pure" monte carlo methods (a stroke of genius). of course i'm biased *drumroll* because i've totally fallen in love with the elegance and power of the method, but it can't be denied that it confers significant benefits in difficult sampling situations.
so, whether or not mlt is being used as a marketing term, the fact is that it forms a solid basis for the current generation of rendering engines, and it's not unreasonable for users to associate "implements mlt" with "efficiently renders almost anything".
MLT has pros and cons, relative to 'vanilla' Randomised Quasi-Monte Carlo sampling (RQMC).
RQMC will converge more quickly for 'easy' scenes, for example a diffuse Cornell box. This is due to the extremely evenly distributed (low discrepency) path sampling made possible by RQMC.
However, RQMC can't compete with MLT (or more precisely, Metropolis-Hastings sampling), for 'difficult' scenes.
Such difficult scenes include the infamous sunlight-through-glass problem, caustics seen in a reflection, caustics on a pool bottom seen from outside the pool, etc...
RQMC will converge more quickly for 'easy' scenes, for example a diffuse Cornell box. This is due to the extremely evenly distributed (low discrepency) path sampling made possible by RQMC.
However, RQMC can't compete with MLT (or more precisely, Metropolis-Hastings sampling), for 'difficult' scenes.
Such difficult scenes include the infamous sunlight-through-glass problem, caustics seen in a reflection, caustics on a pool bottom seen from outside the pool, etc...
yup, and as veach notes in his bible-thesis, if we use only one method then we should choose the more robust one. 20-30% slower convergence in simple scenes is a tiny price to pay for seamlessly handling brutal sampling situations.
this is why commercial rendering engines might like to use it, and why it should be on by default in indigo distributions ;)
and that's my 2c :)
this is why commercial rendering engines might like to use it, and why it should be on by default in indigo distributions ;)
and that's my 2c :)
Who is online
Users browsing this forum: No registered users and 27 guests