Depth of Field

https://inconvergent.net/2019/depth-of-field/

by ingve • 7 years ago

129 22 7 years ago

Why is the randomization necessary?

you get unwanted artefacts in the results if you don't use some randomness. having said that, there might ways of reducing the need for it (which would be more efficient).

dahart3• 7 years ago

It's not necessary, but it's convenient. What's actually happening here is Monte Carlo integration. Randomness guarantees you get the right answer. It prevents correlation artifacts. It also gives you a nice film grain look when you don't have the time to take samples until complete convergence.

tylermw• 7 years ago

Nice. I created a package to generate similar depth of field images in R, using a depth map/image pair. Also offers customizable bokeh shape and some other nice features.

Github: https://github.com/tylermorganwall/rayfocus Blogpost: https://www.tylermw.com/portrait-mode-data/

tomxor• 7 years ago

So pretty! I'm going to assimilate this purely for the aesthetic pleasure :D

ratsimihah• 7 years ago

I know this is so hot! And written in Lisp, I'm having a nerdgasm!

jv22222• 7 years ago

From his generative collection these ones really stand out:

https://img.inconvergent.net/img/gen/20170523-193637-305712-...

https://img.inconvergent.net/img/gen/20170520-230701-136920-...

It's almost hard to believe they are computer generated. They just seem so organic.

sq_5• 7 years ago

It looks like the links you posted are broken because the site owner doesn’t allow hotlinking to those images.

Could you post the link to the blog post with those images? I’d absolutely love to see them!

inconvergent6• 7 years ago

https://img.inconvergent.net/generative/78b7266.html

https://img.inconvergent.net/generative/50519fc.html

Kuraj• 7 years ago

Hitting enter in the address bar allows you to access them

dahart3• 7 years ago

Since the theory isn't really described here, anyone who wants a tad more might look at the Ray Tracing In One Weekend booklet for a touch of the optics math behind depth of field. Chapter 11 derives a defocus blur. http://www.realtimerendering.com/raytracing/Ray%20Tracing%20...

find the new position, w = v + rndSphere(r).

For anyone wanting control over the look of the defocus blur, also known as the "bokeh", random sphere gives you more samples toward the center than the edges. For a circular aperture, it is slightly better & more correct to sample a disk than to sample a sphere, and slightly faster to converge as well. You can optionally use a hexagon or octagon on the disk if you want to get the look of a film-camera type aperture.

inconvergent6• 7 years ago

i would say that what is correct depends on the intention

dahart3• 7 years ago

I'm talking about correctness in the physical optics/lens sense. With lenses, light always passes uniformly through the aperture, where sphere sampling is non-uniform.

Intention is totally fine. This is reasonable if you actually intended to do something different than what a camera does, or if the intention is not physical correctness. This blog post seems to be intending to do something easy for picking samples, as opposed to something optically correct. I'm all for easy, but I also think it never hurts to understand the tradeoff you're choosing, nor to present the harder alternatives.

It's also worth considering disk sampling rather than sphere sampling, because it's barely any harder, and it will make the code converge to the same quality something like 2x faster. Sphere sampling spends too much time in the middle and not enough at the edges. Disk sampling only takes a teeny tiny bit more arithmetic. Jittering & QMC methods will also help a lot with efficiency.

londons_explore2• 7 years ago

This could be made a lot clearer with an animated version either changing the focus depth or rotating the view...

Depth of Field

Related Stories