The Cornell Box tutorial Part I

IMPORTANT NOTE: The renders in this tutorial have all been corrected with a Gamma of 2.2.

So here is the Cornell Box in all its glory:

And this is the scene we are going to use for the tutorial.

The Cornell Box is probably the oldest scene used for comparing radiosity rendering engines. The most popular configuration is described at the Cornell University Program of Computer Graphics Cornell Box data page.

For this tutorial, I invite you to download the A:M Cornell Box project. This is the actual Cornell Box data I modeled in A:M from the published specs for both dimensions and surface properties.

Here are a few things to note about this scene:

First thing to note is this scene is small. Very small compared to a normal room size. The room itself is 55cm wide by 55 cm high by 57 cm deep.


It is important to note that the Cornell Box is small relative to a normal room because room size have a major influence on the way Photon Mapping properties should be set.

Second thing to note is the light arrangement.


The cornell Box is illuminated by a signle flat diffuse luminaire. To simulate that, I arranged 25 klieg lights in a 4x5 array. This is actually one klieg light instanced 25 times in the choreography.

Third thing to note is how the klieg light is setup. To simulate the way a flat diffuse light illuminates a scene, the klieg's Cone Angle is 180° and the Width Softness is 100%. The 100% Width Softness ensures that illumination falls-off at a cosine rate according to the angle from the light surface. This is how light from a flat surface behaves in reality.

Fourth thing to note is the klieg's Fall-Off distance.


The Fall-Off distance should be set is such a way that it doesn't touch any object in the scene. In this scene, you can see that the Fall-Off stops short of touching the walls and the tall bloc. This is especially important when the light intensity is small to ensure that all objects will be directly illuminated in a natural way. The Fall-Off distance could be shorter but ideally not touch any objects in a scene. Sometime this is not really achievable. When this is not possible, then there are other rules to take care of.

Fifth thing to note is the low light intensity.

This is a Raytrace render of the Cornell Box.

For Raytrace render, you would probably want to have higher light intensity. But for Radiosity render, because the surfaces will also contribute light in the scene. light intensity are lower. In fact, light intensity can really be adjusted by trial and error in Radiosity render mode. But there are ways to do quick preview radiosity renders.



OK. You probably want load the Cornell Box project in A:M, go in the Choreography properties pannel, turn ON the Radiosity property and render. I invite to do so and render without multi-pass, with multi-pass set to 1 pass and set to 9 passes to see what you get. I will be back later with more material for this tutorial.

I hope this will be an interactive class type of tutorial so in the meantime, you are invited to post your observations and questions here and I shall answer your questions.

Have fun.

If you did a radiosity render with the default Photon Mapping properties, you should have a result similar to this:


Render Times

And if you rendered the scene with Antialiasing (no multipass) or multipass with 1 pass or multipass with 9 passes, you got basicaly the same result. But the exercise was not futile. There is one important concept that is demonstrated with this exercise. On my computer, the render times were almost the same for all type of render:

10m35s for antialiasing render
06m33s for one pass render
10m55s for 9 passes render

You may have expectred the 9 passes render to take approximately 9 times longer than the one pass render. But for Photon Mapping, this is not the case. For the Photon Mapping rendering engine, the most time and CPU consuming step is the Final Gathering step. And the time it takes is directly influenced by the number of Final Gathering samples. However, the Final Gathering samples are distributed among the passes. So when rendering only one pass, the final gathering need to sample all the 150 samples per pixel while when rendering 9 passes, each pass samples only 150/9 = 17 samples per pass.

So because of that, when you use Photon Mapping, you don't have to worry about using multipass. This will not have a huge impact on render time.

Getting the radiosity render right

The radiosity default settings are rarely appropriate for a given scene. A noisy render like that is indicative that there either is not enough photons in the scene or the Sample Area is too small. Since this scene is small compared to a normal room size, the noise is not so bad. But for a normal room size, the noise can look really bad.

Since increasing the number of photons in the scene will increase the render time, it is always advisable to try increasing the sample area first.

There are basicaly two ways to find the proper sample area. A visual procedure and a mathematical estimate.

I have implemented the mathematical estimate in an Excel spreadsheet:

You may download it here.

So enter the room size whivh is 55x57x55cm and an estimate of additional object surfaces in the room. In the Cornell Box case, the estimate is simple since we know the tall and short blocs sizes which are 34x17x17cm and 17x17x17cm respectively. And since the room itself is missing the front wall, a 55x55cm area is subtracted which gives 443cm additional.

Simply enter the suggested Sample Area value in the property box and rerender at 9 passes. You should get a result similar to this one:


Now that you have this Excel utility, you may want to try with lower and higher number of Photons. See how low you can go and still have acceptabe radiosity renders by setting the appropriate sample area. And see how the number of photons in a scene affects the render time.

Next time I will review the visual procedure for determining the appropriate sample area. This procedure can be usefull for situations where it is difficult to measure or even estimate the total surface areas in the scene.

Have fun.

Sorry Yves, It had to be done

BTW, does motion blur work with Radiosity?

eric

Hi! I'm new with the lighting process, and especially with radiosity in A:M (I never use it ) So I'm very excited to read all the information you can bring us ypoissant, and of course all the tutorials you can make

Well, I made the render for this test scene, and the render times is about 25 minutes (well, I haven't a very powerful computer)

Then, I wanna try to do a little animation with the same scene and see how works radiosity. Then I animated 21 frames of a bouncing ball, and the render time was about 3h. 16 minutes at a resolution of 320x240.

But the problem I see is that the light or the image is constant flickering. This is normal with radiosity? Somebody has tried to make an animation?

I put "on" the multipass to 16 passes and motion blur to 40%

Cheers.

Awesome animation!! I havent seen anyone do an animation with radiosity yet, and it looks so realistic. I'm not sure why its flickering, but that looked great!!!

Hi Yves,

I was just curious to why the cornell box only has 3 walls? (besides for camera).
If I,m understanding this correctly(photons are shot from light and bounce around room) its seems like a lot of the photons would be lost
(escaping the room)

Which brings me to my next question how long do photons bounce around? Does each one bounce around the scene hundreds or thousands of times.

Would you be able to use a substantially lower number of photons in a closed environment to achieve same results?

I think you set the number of times they bounce around. Default is like 15 bounces.

Render Time: around 20 min.

The black thing is like.... a ghost thing

Ummmmm. How/where do I select radiosity render and photon map settings? (I must be blind - using v10.5r, 11.0t or 11.1)

Cheers

Never mind... Found it in the online tech Reference. I have Will Suttons v11 Tech Ref but the index really sucks!

Cheers

Thanks MATrickz,

For some reason, I was thinking you could not control the number of bounces

My old Cornell Box in Brooklyn was bigger, but this one has a view of the river and the rent is only $800 more a month.



Render time 31 min.

My Q's:

-AFAIK, I did this with the default settings but it didn't come out nearly as "spotty" as the "default" setting render near the top of this thread. Likewise with the unmodified Box project with no character in it. So what changed?

edit:It seems that the sample project already has 750 set for the "sample area" which is the number the calculator suggested. So... the "Spotty" render above must have had a much smaller number?

-Would more RAM help these renders, or do they use pretty much the same as regular renders?

I think , you should increase the sample area, because the value for your additional surface is larger as in the sample project.
The calculation for the additional surface is only easy in the sample project, in your case is it not easy , while first you must calculate the surface for all objects in the box, then you must substract the surface , that is in contact with the roomsurface (doubled , objectsurface and roomsurface). In the sample project this are the bottom surfaces from the two boxes .
For the sample project the calculation for the additional surface is as follow
Box1 34x17x17

Good information. ...and nice pics too! Wow.

Robcat,
If only you had been about a month earlier on that design... that (just slightly adjusted for location) would have made an awesome A:M CD image.

Just reading-up late on this thread, busy weekend...

In Ernesto's test, shouldn't the ball have bounced light from the source back-up to the ceiling and lit it up a bit? (As it passed close by)

wondering...

Cornell Tutorial part III
Finding Photon Properties visually

The alternative way to determine optimal Photon Mapping properties is to directly visualize the Photon Map. To demonstrate that, we will use the Original Cornell Box project.

First reset all Photon Mapping properties to their default values. Number of photons : 10,000, Sample Area : 100, Photon Samples : 100, Max Bounces : 15.

Then set Final Gathering to OFF.

In the render to file pannel, set multipass to ON but set the number of passes to 1.

A render should produce something like this:


This is the polka dot syndrome. When the sample area is too small, you can clearly see the polka dots. The isolated polka dot size is exaclty the size of the sampling area. Ultimately, you could see each individual photons if you rendered large enough and with a sample area small enough. Try setting the sample area to 20 and render for an example.

The idea of the visual procedure is to increase the Sample Area untill the direct visualization of the Photon Map gives almost acceptable renders. For instance increasing the Sample Area to 250 gives this:

which is still not quite acceptable.

But increasing the Sample Area to 750 as computed by the Excel utility gives this:

Which would almost be usable as is.

When you reach a point where increasing the Sample Area does not improve the smoothness of the photon map anymore, then you have reached the limit of what the Photon Samples can give you. If you want to improve a little bit further, you will have to increase the Photon Samples and then the Sample Area.

For instance, increasing the Photon Samples to 500 and then the Sample Area to 1700, you would get this:

which, I personally think is pushing it a little too far as most of the irradiance gradients are completely lost. There is a good compromise to be found somewhere.

Note

Direct visualization of the Photon map is much faster than with Final Gathering which makes it suitable for finding appropriate properties by trial and error. Because of that, it is tempting to try to find direct visualization properties that would produce perfect illumination. However, it will not be possible to get a perfect illumination in that way. This is intrinsicaly related to the way irradiance is computed. So be already advised.

The interesting aspect with this method of searching for the optimal properties is that you have a visual way to judge the irradiance distribution. It is a good practice to use both methods. That is finding good estimates with the Excel utility and improving the estimate with the visual method.

Keep those renders coming.

First of all, i have to thank you for this tutorial, i´ve spent hours and hours without any good result. Now, there is hope.
I did some rendering with the cornell box, using your excel sheet.
From left to right:
Photons 1000 sample 2450 (1 Pass) = 52 sec (9 Pass) = 1:29 sec
Photons 10000 sample 750 (1 Pass) = 56 sec (9 Pass) = 1:38 sec
Photons 20000 sample 550 (1 Pass) = 58 sec (9 Pass) = 1:46 sec

looking at the pictures i think you can say: The more Photons you use, the brighter the colors are, but i think if you use to much Photons the picture can be unrealistic because the objects seems to glow from inside, the reflected light seems to be brighter as they should be.
Am i right???
Greetings
Frank

The way to help control the "glowing" objects when you increase the number of photons is to increase the photon samples and then the sample area.

Just thinking, can´t you use the glowing to show more informations? If you make a summer picture with realy strong energy light you show the glow, otherwise you control it like you said?
Greetings,
Frank

Even though Radiosity is designed to emulate real life lighting, the controls you have still leave place for creative tweakings. You could certainly find creative use of those settings.

Very fascinating! Thank you so much for this extremely useful information! I will try to apply this to some figure renders.

to Yves:

You could certainly use more or less or even only one klieg light. Whatever suits you.

In my case, I was trying to replicate the original Cornell Box as closely as possible for testing purposes. And since the luminaire was a rectangular one, fitting a round klieg in a rectangle was not optimal. I could have used more than 20 but I thought 20 was close enough.

Bonjour Yves,

Thanks for a real interesting thread.

A lot of great posts are coming out of this as I think these types of images are what many of us in the AM Community are striving for. Certainly, it is a testament to what AM is capable of and how little most of us know of this program.

Anyway, I wanted to know how one should approach a project from the standpoint of Radiosity. Should I model my entire scene first and then throw a huge box around it for Radiosity to work?

Any advice would be appreciated,

Eugene

Wow I love this new thread. I'm trying the tutorial right now cant wait to see the results of my rendering , I just have one question, what is the difference of multipass with 1 pass and multipass with more than one pass, i'm actually not sure what multipass is except that you get antialiasing without it on.

With multipass, you set the number of renders that are averaged together to get a final image. Each image is slightly offset from one another which produces antialiasing. And the light is also sampled at different positions on its surface area to produce soft shadows.

When you visualize the photon map directly, you want only one pass in order to clearly see how the photons cover the surfaces. If you used more than one pass, because the light is sampled at different positions at each pass, it would be difficult to figure the actual area covered by the photons.

question:
are all features of a:m supported by photon mapping?
I tried a translurenz plane in front of the light and it doesn´t look like i´d thought. What about volumen light and transparent elements?
Greetings
Frank

Those features should work.

It is very difficult to answer an abstract question. Please post images, examples and renders so I can comment / feedback / advice in a more specific way.

Without multipass, antialiasing is done with another technique called a-buffer.

Cornell Box tutorial part III

Computing the Photon Samples with a calculator

The idea is to estimate a sampling area that will most probably leave no holes in the photon coverage. The information we need for that are : the total surface area to cover with photons and the area that is covered by each photon irradiance estimate.

The most significant measures to help estimate the surface area to cover with photon is the size of the room. Here, the room was 55x57x55 cm. So the total surface area of the room is the sum of each wall areas. There are 4 walls of 55x57 and one wall of 55x55 (the front wall is missing) so we have (55x57x4) + (55x55) = 15565cm². Then we need to add the estimated area of all the objects in the room. In most situation, it is not practical to actually compute the surface area of each objects the way we did it for the Cornell Box. A rough guess will do just fine and an overestimated guess is better than an underestimated guess. So let's say we guess 5000 additional cm² and let's round the total estimated surface area to 21000cm².

The next measure we need is the coverage of each photon irradiance estimate. We shoot 10000 photon on 21000cm² surface. That means that we will get approximately 1 photon per 2.1 cm². And since we use 100 photon samples for each irradiance estimate, each photon irradiance estimate will cover an average of 210 cm².

With this estimated photon irradiance coverage, we can estimate the sample size in this way : The radius of a circle covering a 210 cm² area is SquareRoot(210/Pi) which gives 8.18 cm. And since the sample area is given in 1/100th of cm, we get get a sample area of 818.

Recap

Given a room where W is the width, H is the height and D is the depth,

Scene Area = (W*H*2) + (W*D*2) + (H*D*2) + Additional object areas.

Irradiance Coverage = Scene Area * Photon Samples / Number of Photons.

Estimated Sample Area = SquareRoot( Irradiance Coverage * Pi ) * 100.

Hello friends of the cornell box,
looking in a book about japanes architekture i found a picture i think you like to see. It´s a realtime cornell box.
Greetings
Frank

Yves
Yesterday I have spend some time to fiddle out, why I get for the surface calculation different results to your's. (the demo project)

The solution was , that the room in your project is 55x56x55 (the blocks have also little different sizes)

Kind Regards
Steffen

Hi Steffen,
is this a new plugin your working at? I´ve just looked but can´t find a "calcsurface". It would be a good helper for calculating for Radiocity.
Greetings,
Frank

Steffen,

Yes, you are right, My measures were taken from visual inspection. There is no real need to be precise. Also, I took into account the fact that the two blocs are sitting on the floor and thus subtracted top and bottom surface. I also did not include the luminaire because it roughly has the same surface as the one it covers.

This said, we had the idea of adding a button next to the Sample Area that would calculate a suggested sample area from the surfaces of the objects in the scene. It seems you are already working on that

Yves

Hey Steffen thanks for this little plug. I don't have excel on my pc right now and don't deel like loading it .

Now online , I think have some glitches , but You must find them :-), also no documenatation
Download the plugin

Short description
- plugin is only for A:M V11 and V11.1 (PC offcourse)
-Start the plugin from the chor , the chor MUST be in the Frontview (Numpad-2), otherwise you get false results for the total surface .
-You can change all inputfields, but hit after them "Recalculate" to get the correct sample area,
-The changed values are stored in the correspondending propertys in the chor after hitting OK , Cancel return the plugin without any changes .
-Use it at your own risk ....

Kind Regards
Steffen

Ive just read through all 4 pages of this topic, its great stuff.
I dont quite understand it all just yet, but im sure i will eventually.

I downloaded the sample project and the Exel doc and heres my render of the Cornell Box.