Tag Archives: Siggraph

2.5D relighting in compositing with Blender (using spherical harmonics)


About a year ago, I was watching a great presentation at the Nuke Master Class by Roy Stelzer (2D TD at Double Negative) about “2.5D Relighting inside of Nuke”. There is 3 videos which you can download from The Foundry and also files which goes with. I recommend you to watch those videos especially if you are not familiar with relighting in post-production.

What Roy did, is using a paper from Siggraph 2001 called “An Efficient Representation For Irradiance Environment Maps“.

“An Efficient Representation For Irradiance Environment Maps”

This paper cover how to do Environment Lighting based on Angular Map (or light probes, or HDRI map, whatever you want to call it) but without using those map for computation. Odd you said ? Yes sure !
They develop a technique using Spherical Harmonics to simplify an HDR Angular Map to only 9 coefficients !! That means after their process, you only have to do a few multiplication with those 9 coefficients and then you are done !  Check out the comparison  :

left is a standard rendering, right is using the 9 parameters. Grace Cathedral Light Probe was used for this rendering http://www.debevec.org/Probes/

If I’m not mistaking, there is less than 1% of error between both rendering, and of course you can image that the second one goes much much faster 🙂

Please visit there website and presentation for more information about it :

What did Roy ?

In his demo, Roy shows the use of  the irradiance only of a Angular Map. This means he will get all the HDR Luma intensity of the map (no color variation). Once you have the 9 coefficient, the operation is pretty easy and so it compute very fast. The operation looks something like this :

color = c1 * L22 * (x2-y2) + c3 * L20 * z2 + c4 * L00 – c5*L20 + 2 * c1 * (L2_2 * xy + L21 * xz + L2_1 * yz) + 2 * c2 * (L11 * x + L1_1 * y + L10 * z)

Where c1, c2, c3, c4 and c5 are 5 constant number given in the paper. X, Y, Z are float numbers from your normal pass (X2 means x*x ; xy means x*y ; so on … ). All the L22, L20, …. variables are the 9 coefficients)

So as you can see, this is not a really complicated operation and it does compute really fast. Running this on each pixels would return a kind of light map (irradiance) which you can use to ADD to your original color map.

What did I do ?

As I did for the Lift/Gamma/Gain the first time, I tried to reproduce this formula with the “Math Node” of Blender. So for that matter I did use the EXR file (render passes) provided by Roy in his demo, and only kept the Normal and AO passes.

Blend File here

This is not as fast as it could be, the render time for a 1920×1080 is around 1.5 second (well for HDR environment lighting we have seen worst ^^). There is several reasons for this to be slow, but I’ll come to that later.
Note that for this example, I did use the Grace Cathedral Light Probe value and not Roy’s light probe.

I was kind of happy of the rendering though, but a bit disappointed to only get Luma value when environment maps have so much information about colors as well ! (you thought the previous example was a mess with nodes, wait for the following one 😉 )
UPDATE : I was totally wrong 🙂 !!! the only reason why I only get Luma (or actually greyscale) is because I used Math node. I thought it was able to do the operation on any kind of input, but actually it does it on only one composant. So the vector operation never happen :p. I just figure it was possible by trying this same formula in another shader language (pixel bender) and see color happening ^^ . So my bad, Color works too, and I’m not sure to know the difference between the vectors or the Matrices in this case, except using the formula with vectors is much faster ! (I’ll change the blend file later)

So I took a closer eye to the paper, and especially to the example they provide, and I found out that their filter wasn’t only generated coefficients, but also Matrices !!! This means you can do the operation with 9 coefficients to just get the irradiance of the environment or do a similar (but a bit more complex) operation with 3 4×4 matrices (red,green,blue) !
I guess the obvious reason Roy didn’t go for this solution was because the computation is more slower, and he didn’t really need it since he is doing a kind of 3 points lighting in his example.

As I said the math are a bit different ! Here is the formula by using the 3 matrices :

n = worldNormal;
color.red = dot(n , matriceRed);
color.green = dot(n , matriceGreen);
color.blue = dot(n , matriceBlue);

Ok so while this might look more simpler on paper, remember the matrices are 4×4 and if a dot product is quite simple, it is not the costless operation too :). Here is what it looks like in Blender with “Math node” as well :

Blend File here
Again, due to the heavily use of Math node, I believe that 3 second is not too bad, but I’ll come back to that later. Also the node I use to rotate the normal pass, is using some math that might slow the render a bit too and not absolutely necessary

This shows that the technique is working pretty well, but probably not production ready as it is in Blender since we are missing a few thing to make this work smoothly.

What would we need in Blender for this to be more efficient ?

  • More “Input” nodes : this might be one of the reasons why the rendering is slow down a bit. Because actual “Input node” only works between 0-1, and the matrix number were between -n and n, I had to find a trick. For this I used the “Math node” for each number of a matrix. Setting it to “Add”, enter the value in the first input, and setting the second input to 0.0. So the output would be equal to the first input. I only did that because I couldn’t figure another way to have negative and positive values in Blender compositing in another way. But this also mean that the compositor tell blender to do an operation for each input, can’t say it’s much optimized 🙂
    • Input Value that goes between -n and n
    • Vector4 Input (right now you can only use Vector3, but you usually need to work with alpha)
    • Matrix Input (could even be useful to do some quick and dirty convolution)
  • Expression node : ok now, I’m dreaming for this one ! And this is IMO probably the main reason why this is so slow. I believe that each time I’m using a Math node, Blender does treat them individually. Which makes sense though, but it probably make a lot of exchange with inputs, outputs, inputs again, outputs again, ….
    I would believe that sending the entire operation at once to the CPU (or whatever) and getting it back at once would make things different and much faster (but I might be wrong on this one !?)
    Anyway the other reason for this node would be … well seriously, have you seen the mess with all the nodes ?!?
    So a simple field, even just interpreted by python would be great !!!!
  • Math with Vector : Maybe I did something wrong, but I couldn’t do a “dot product” between vectors, which is one of the reason why I have all those nodes. I’m doing the entire “dot product” by hand and this is heavy.
    I wish Math could be use with any kind of input. But again, maybe I’m doing something wrong here
  • Passes : we need more passes! For this example we need an Object Normal Pass rather than a World Normal Pass. Probably not to much to do though, the only problem I have with the Passes system today is that they are all hardcoded in Blender, which makes it complicated to create a custom one like you would have in Maya.
    I’d like to be able to assign a material overall to a define passes, but yet I would probably need to write shaders, which implicate the needs of writing shaders as well for the renderer. I guess OSL will fix that in a future if it gets implemented one day
  • Better support for EXR : beside this really annoying, flip node you have to add when working with other packages (I know flip node is nothing, but when working with 2K or 4K it is not the same deal, especially when the composite gets complex, you want to save any operation you could) , but I believe anyone is aware of this now, the other lack of EXR in Blender is the passes support. it doesn’t support custom passes coming from other package. Roy provided his Nuke script, with all the EXR file so you could play with it. But when I tried to load it in Blender, the input node couldn’t find all the passes inside, beside the usual color, Z (and maybe another one can’t remember exactly) it couldn’t find any. So I had to open the EXR in Djv, select the pass I wanted, and save it to another file as the RGB value. Really painful process

Hopefully someone hear me out there ^^

Blender 2.5 expected in October 2009 ?


The first release of the 2.5x series is expected to be available in october 2009

This is what you can see at the bottom of the Blender Siggraph 2009 Flyer which you can download on Blender.org here : http://download.blender.org/institute/25_flyer5.pdf

So now we get official annoucement for the following features :

  • Still free (no kidding :p)
  • New Animation system
  • New interface
  • Macros, History and Search feature
  • New Python Scripting API
  • Custom Key Binding (let you make your own shortcut … swweeeeeeeetttt)

The entire Blender Siggraph Press Kit can be found here

RollingShutter Plugin by the Foundry for After FX and Nuke Released !!!

I talked about the Foundry working on solving the rolling shutter issue on a previous post. It seams they were waiting the Siggraph to announce it. At the time I though it was only for Nuke, but the great news is they also port their plugin to After Effect CS3 and CS4.

It seems they fixed the horizontal panning issue, but still not the vertical one, which at this point looks impossible but I’m actually waiting for another paper on stabilization which should show up at siggraph too. Anyway it’s going to help a lot for matchmove !!!

Official RollingShutter by Foundry plugin page