MLT - how difficult is it?

[Kuros]

I've implemented path tracing, light tracing, bidirectional path tracing with MIS, and they all give the same output for any indata that supports each. I check this by flipping the screendumps in mspaint (ctrl-v / ctrl-z), which makes me notice immediately if anything is even slightly off. When I do this with the comparison screenshots from many people who have implemented MLT (even Veach himself) I notice that they haven't managed to get MLT implemented completely correctly.

So I'm wondering, how difficult is it to implement and get Veach-styled MLT right compared to, say, BDPT with MIS? How much knowledge do I need on MCMC and Metropolis Sampling etc (it's a new area for me) to be able to fill out all the gaps not mentioned by various paper authors on the subject?

[Kuros]

Okay, how about Kelemen style MLT then?

[spectral]

It is like everything here...

First, you have to read the related paper... in your place I will take a look at Kelemen style... it is not too difficult if you understand everything !
It just request some time to debug it... by example, take a look at the LuxRender code or Mitsuba.

[ingenious]

I check this by flipping the screendumps in mspaint (ctrl-v / ctrl-z), which makes me notice immediately if anything is even slightly off.

That's also my usual method of validation, however it's good to dump HDR images from time to time and compare their difference. Some inconsistencies are difficult to spot in Paint.

When I do this with the comparison screenshots from many people who have implemented MLT (even Veach himself) I notice that they haven't managed to get MLT implemented completely correctly.

And what exactly do you mean by that?

[Kuros]

I check this by flipping the screendumps in mspaint (ctrl-v / ctrl-z), which makes me notice immediately if anything is even slightly off.

That's also my usual method of validation, however it's good to dump HDR images from time to time and compare their difference. Some inconsistencies are difficult to spot in Paint.

When I do this with the comparison screenshots from many people who have implemented MLT (even Veach himself) I notice that they haven't managed to get MLT implemented completely correctly.

And what exactly do you mean by that?

What I meant was when even the geniuses get different outputs, maybe I shouldn't even try. (the glossy reflection on the floor in the rightmost corner of the teapots on the table picture from Veach's thesis for example)

[ingenious]

What I meant was when even the geniuses get different outputs, maybe I shouldn't even try. (the glossy reflection on the floor in the rightmost corner of the teapots on the table picture from Veach's thesis for example)

Don't jump into conclusions. Metropolis light transport cannot be compared with Paint that easily. It's noise distribution is not uniform, and also there's a global scaling factor that can be off. The images may look identical with a larger number of samples.

[wjakob]

The following may be useful to you: http://www.mitsuba-renderer.org/manifolds/. It contains implementation of all common variants of MLT, including the one by Eric Veach.

[spectral]

Wonderful,

Great works... I'll explore it more

Thanks for notifying it !

[toxie]

And if you wait some more days(?), then there will also be course notes available for this: http://s2012.siggraph.org/attendees/ses ... -synthesis
Part 5 has a neat overview (with some implementation snippets and at least some of the nasty details) over the topic..

[ganqu]

... there will also be course notes available for this: http://s2012.siggraph.org/attendees/ses ... -synthesis ...

Where should we expect this notes to appear? Sorry if this is a sort of a dumb question, but I looked up some siggraph courses pages from previous years and didn't see any materials there. However I found some notes materials on non-siggraph-related webpage (one of the authors personal page) with simple searching the web. Thank you.