Looks good. :)

At the moment, it does seem that you can still see the equations on the internet archive version: https://web.archive.org/web/20130121205251/https://the-witness.net/news/2011/07/irradiance-caching-part-1/

Look at all of these amazing results to this really hard problem you are facing but all the equations are gone! Is there any way to revive the equations?

Here is a link to a page in a book. I was listening to a reading the other day by its author, Osho, and thought of this game. I can only hope there is some relation. Even if not, I’m sure the curious player will still be able to create their own interpretation.

Justin – part of me feels that even if the game hasn’t any direct storytelling in it, there will be loads of implicit story and meaning in every piece of the puzzle. The Witness as a title reminds me of buddhism – buddha meaning in many ways “the witness.” The awakened one who observes. It is very much part of zen philosophy to be a silent observer, and that may be the player’s role in this game. I wouldn’t be surprised if it were up to us to infer meaning on our own… and if so, I think it will make the game so much more amazing. :)

http://www.youtube.com/watch?v=00gAbgBu8R4&feature=channel_video_title

Ignacio, Jonathan… is this for real? if so why isn’t anyone talking about it or investing or trying to replicate it? Ignacio worked for Nvidia, right? are those guys looking into anything similar?

a year ago they showed and earlier demo and they got some attention. mostly from people that said it was not true or just a joke or whatever, but why would they come back? I mean if nothing can run this why bother?

http://www.youtube.com/watch?v=Q-ATtrImCx4 (year old demo)

Will links to articles be posted here… i can never find out about indie games coverage. N4G doesn’t like indie games and a lot of hub news site never post previews on indies, but if i found news of the game from other sites i will not only share them here but i will try to publish it at N4G and other hubs.

Also, will The Witness have a story that is heavy on Science and Science Fiction like Braid was?
I love the story of Braid and it thought me the importance of science in the world and the importance of getting serious about the objective world, but at the same time looking at things from many angles from a subjective perspective.

Like making an atomic bomb or going to space, i hope the story is about deep and meaningful things to the human race. i hope , like Braid can get me even more interested in science!

Also… who is the voice actor? is it the same guy that did the tutorial from Raspberry, he had a charming voice! is he your friend or something?

Is this used often? do other games have it? is it popular or used in AAA games or no other game has this, is this revolutionary, will it help games get easier to develop in the future… ?

i don’t know if this is new for real time graphics, but if its not… you guys have ton of friends in the industry. why not just ask for help? i don’t know about the ethics of game development or if you can use what someone told you but maybe you just don’t believe that other games use it right? so you have to go through all this yourself…

That would be a cool post to see! “Ethics of Game Design and Asking for Help”

That shot is actually using our default settings and corresponds to the lightmap view at the end of the post, which I think does not have obvious artifacts. In that shot the minimum distance between samples is limited by the lightmap resolution. The samples that are farther apart, are so, because they are far enough from occluders, and because the irradiance is not expected to change abruptly. I basically tweaked our error thresholds to get the desired visual results at the minimum cost and that results in the sample distribution depicted in the screenshot.

With our high quality/production settings, artifacts are gone entirely, and at the medium/default settings they are rarely noticeable. Note, that our results are biased, we trade accuracy/correctness for smoothness/good appearance. In fact, some of the scenes that looked noisy without caching (due to the limited hemicube resolution), looked much better with caching enabled.

Ok, fair enough. In the sample shot at the top the sample points seem pretty far apart, but maybe that was captured with the same “exagerrated to show you the effect” settings.

I guess another question is to what degree the look with correctly tuned settings has any visible mottling which isn’t actually present if you don’t do caching. I’m not experienced enough with radiosity to know what things “should” look like. Is it tuned so there’s no visible effect? Or is it subtly visible but acceptable?

Thanks! I updated the link.

fwiw– the siggraph course link is broken … looks like it’s moved to here: http://cgg.mff.cuni.cz/~jaroslav/papers/2008-irradiance_caching_class/index.htm

Yeah, Jon also suggested I added a section explaining how the gradients are used, I thought it was a good idea as it would clarify things a bit, but our implementation is completely standard in that regard, and time for my #AltDevBlogADay post was running out (in fact I was 1.5 hours late), so that got cut out.

I think you have a good point, but you have to keep in mind that this whole thing is a very crude approximation anyway. As you translate your sampling point new radiance sources that were previously occluded could come into view. The same thing happens with rotations, sources below the horizon may become visible under rotation. These effects are largely ignored, but despite of that the approximation provides better results than using no approximation at all, and it does so at a very small cost.

Keep in mind that the approximations are only OK with very small rotations and translations. One of the things that are also estimated when analyzing the hemicubes is how accurate are the approximation expected to be, and this is used to guide the record placement.

Sometimes the estimates fail miserably and that results in artifacts. This usually happens under translations, when a bright radiance source becomes suddenly visible. Existing methods don’t deal very well with this scenario and I think I came up with a innovative solution to the problem. Rotations are usually not an issue, because the radiance samples are cosine weighted, so the new samples that appear near the horizon have near-zero weight.

I personally think there’s still quite a bit of room for innovation on this area, but I pretty much limited myself to adapt the existing methods to our setting and work around the problems that we found. Once our implementation was working well enough I just moved on, since we have many other things to do that have a much bigger impact on the final game.

Anyway, ignoring rotation, I assume what you’re computing with these gradients is a linear extrapolation, something like: E(X+dx) = E(X) + u*dEdu + v*dEdv. Whatever this formula turns out to be, I would think the task isn’t actually to compute infinitesimal gradients at X, but to compute the optimal numbers that supposedly are “dEdu” and “dEdv” that allow you to extrapolate as far as possible (e.g. in a least squares sense). One could hypothesize some sort of higher-order extrapolation as well, and it might be more tractable under this model where you’re not looking at gradients, but looking at it as generating approximating polynomials or some such. However I don’t know what the math would look like for trying to model things this way. It just seems a weird thing to me to start off with the gradients. (I guess this is a basic well-known thing; the gradient is a particular truncated Taylor series to the actual function. But truncating a Taylor series doesn’t actually give the best approximation for that order.)

I’m not sure how gradient rotation can work when you have a hemisphere(cube) of samples instead of a sphere of samples. Is this only used for small rotations (e.g. on a curved surface) and not at large rotations (e.g. a 30-degree joint)? Is it only used at internal angles (where visibility means less than a hemisphere is actually visible anyway, so nothing is missing) and not at external angles (where there can be missed samples “around the corner”)?

Temporal stability is not a great concern for us, because our lightmaps are static. In some areas we have multiple lightmaps for different lighting configurations and simply blend between then. In that case Irradiance Caching could introduce some stability artifacts, but that is not very noticeable as long as the quality of the lightmaps is high enough.

You already be aware of this, and it doesn’t really affect your lightmap baker, but I believe one problem with irradiance caching is temporal stability. If you render multiple frames with moving lights you tend to get popping as different adaptation decisions are made. I think there are alternatives for this, but I forget the references. Anyway, just thought I’d mention it.

ta,
Sam