END OF AN ERA, FRACTALFORUMS.COM IS CONTINUED ON FRACTALFORUMS.ORG

I'm using Mandelbulber. I turn on the depth of field post-processing, and it always looks wrong. I finally figured out why.

In actual lens blur, light wins over dark every time. Take a picture of a perfect checkerboard pattern of black/white, blurry, with a camera. The white squares bleed into the black, never the other way around, but Mandelbulber (and I assume other depth of field algorithms) just use gaussian blurring, whereby dark colours are simply smeared with light ones.

If before blurring, you apply a convolution matrix of positive numbers, eg:

[
0.1 0.3 0.1
0.3 0.5 0.3
0.1 0.3 0.1
]

Then the entire image will be "light colours propogated" and if you then blur that, you will get something more closely approximating lens blur. It's still very imperfect, but should be more natural-looking than the gaussian-only method.

I tested this out by pulling up GIMP and doing a generic convolve on my image before blurring. In that case, it's a 5x5 matrix rather than 3x3 as I've illustrated here, but the principle is the same: you put the highest value in the centre, and as you increase distance from centre of matrix, the values are proportionally smaller (probably on a quadratic basis, but I haven't completed my analysis that far yet).

I think to do this 100% properly, you'd need the convolve matrix to be larger the further from the focal plane the voxel is. But again, as a first approximation, just doing it with a static matrix would be an improvement over what's there currently.

Does anyone have opinions on this matter?

90% of the time, users apply standard gaussian and box blurs to simulate depth of field. This couldn't be more incorrect. As a result, I spent a lot of time studying correct depth of field and them creating a Photoshop plugin that does it correctly.

You should check out the technical info as it's interesting, especially when you break down what a correct bokeh looks like:

http://richardrosenman.com/shop/dof-pro/

Specular bloom, as you point it out, is something different though and coincidentally enough, I created a plugin for this too: http://www.lumierefilter.com

Cheers,
-Rich

Hi Kram;

That's a really good and important question. The order absolutely matters. For instance, in my field, we have to deal with motion blur and depth of field regularly. Rendering these in 3D is computationally exhaustive so we often rely on applying them as a post process. However, adding motion blur first and then depth of field yields incorrect results. Likewise, adding depth of field first and then motion blur is also incorrect. The bottom line is they both happen concurrently in the camera yet there's no tools out there that allow you to do both in one shot as a post process.

Often enough, we can cheat it and you won't notice but the order is definitely an ongoing headache we deal with every day.

With phenomena like blooms, I think it's easier to cheat because a bloom is equivalent to a gel or filter over the lens. As a result, it's ok to add motion blur and/or depth of field first and then add the bloom. Likewise for vignettes. At least, that's my opinion.

-Rich

That's why, as computers get faster, physically correct renderers are increasingly attractive: human effort isn't getting any cheaper, but FLOPs most certainly are...

Also, while you're taking extra samples, it doesn't really cost any more to incorporate depth of field, motion blur, area lights, etc. In other words, it's not like it takes twice as long to do DOF and motion blur, compared to just DOF alone. So that extra sampling effort to do things correctly is quite justifiable, since it covers all effects if you do it right.

Exactly. BUT, and there's always a but. Motion blur, dof, etc, is free in unbiased renderers (Maxwell, Octane, Arion, etc) and unfortunately, those are still not viable options. They still take waaaaaay too long when compared to biased renderers and those that utilize GPU processing, have some serious limitations which prohibit their use for any serious commercial production. For instance, depending on the project, it is not uncommon to have geometric scenes in excess of millions or billions of triangles. This is immediately impossible with GPUs. Likewise, using 4k textures, or many of them is also a limitations due to GPU ram. There's much more.

So your standard commercial production renderer like VRay, Mental Ray, etc, do cheat and that makes turning on a features like depth of field, quite a bit more taxing. For this reason, I'd say about 80% of all projects still use post depth of field. VRay and Mental Ray do have GPU extensions (VRay RT GPU, IRay) but like I said, they are limited and more useful for single frame production than animation.

It's also interesting to note that film production renderers weren't even raytracers up until recently. Renderman, 3Dlight, from what I know, didn't even do raytracing in order to be able to achieve the speeds required to render at film resolutions quickly and efficiently. I believe films like the Matrix were among the first that used commercial raytracers like Mental Ray to render with global illumination and such.


You probably could but it would be extremely complicated. For instance, you need to supply a depth map which is a frame that describes the distance of the scene to the camera:





With this you can effectively apply and modify the depth of field:



You would have to do the same to the motion blur: supply a vector map that would describe the direction and length of the motion vectors. This exists and you can do it with select apps. So imagine having to supply two frames of depth and motion information, both of which also have to be computationally derived, just to process one frame? And then imagine doing it for a sequence? Nahhhhhh....

-Rich

Richard, I am delighted that you are a member of fractal forums. Reading your postings really ups one's IQ

  Why'd you blur the sky so much?   (joking) 

  So is a velocity map the same thing as a distance map used to taper motion blur on more distant objects?  You could implement a linear map to decrease motion blur with distance- the mountains, farther down the track, the sky....

Hmm.. The train map example seems to only take into account x-y-plane blurring, judging by the velocity map which apparently lacks a blue-channel which, I'd assume, stores z-velocities.
If that's "good enough", why don't you just store the z-depth information in the blue channel and apply DoF effects accordingly?
Or am I incorrect in assuming lack of z-velocities?