Cycles seem to be time consuming.... 9hrs/image!!

I’ve been pounding on this for some time now so I’ll ask it here.

When someone is doing a still image. Taking 9 hours to produce a truly realistic effect is awesome. But when you want to animate. Why spend 9Hrs/frame(image) to produce a realistic look in cycle??? This is crazy.Unless you’re in possession of a hellish render farm. Even the Blender Foundation with their Mango project are going to be hard pressed to produce realistic renderings hare they not??

By the way this is something I hope to hear from them in the Mango DVD.

My point is. I still don’t understand the fuss around Cycles. I’m aware that Cycles is a WIP for a few years thus it will be faster in the future but still.
I’m also aware that I’m part of a minority here for everyone else seem to love cycles so I guess there something that I’m missing here.

So I’ll ask this question:

How do you(anyone) intend to render realistic animations in Cycles without spending months and months of rendering?

There are technologies being developed in the GPGPU area. Also, big productions always had to find alternatives to some special effects, like using billboards instead of 3D models when possible. There are also various composition alternatives that will reduce noise, smooth and perform general adjustments that would cost rendering time.

To put it simple:
1 - rendering full HD photo-realistic animations isn’t for the average budget/set up yet
2 - … and also isn’t for the average user

Where did you get that 9hrs figure? How fast a cycles image converges to an entirely noise-free scene depends foremostly on the lighting setup and then the complexity of the materials used. (and to a lesser degree, scene complexity)
For an animation, you don’t need an entirely noise free image either, because you can use (among other things) temporal noise-reduction filters that re-use samples from previous frames to some degree.
Of course Cycles won’t enable everyone to render long animations at high quality - but pretty much all other renderers don’t do that either.

It’s actually a copie/paste reply I wrote in a Blendernation forum: Half Life 2 Gravity Gun post. He said he spend 9H (3500 passes) for rendering his image.

A couple of things about cycles and speed:

  1. If you make a material and lighting setup that is as complex as you would do in BI, you will probably get a similar speed. But that is not what you want to do with cycles, now do you? Everyone loves to crank up the GI and use realistic lighting, but expects good speeds with that too. Real is slower then fake approximations. You may not be able to do much of the fake approximations in cycles right now, but you will probably be able to do that in the future. Cycles got Ambient Occlusion (which is a fake effect to imitate GI) just a few days ago. That tells a lot.
  2. Cycles shines when you have a decent GPU. With my Nvidia GTX560 I am more or less happy with the performance.And the blender foundation will have no performance problems because they will have a bunch of Ndvidia Teslas (so I’ve read somewhere).
  3. Cycles will be 2-3x times faster in the future, as said by Brecht in the last blender conference.
  4. Cycles will probably have other rendering algorithms in the future, such as photon mapping.

Cycles is not loved for its current feature set or speed. It’s more looked into how much potential it has.
It’s loved because of the magic of finally having a good rendering engine bundled with blender, and the ease of use, nodes, realtime preview… That kind of stuff.

So… you took a figure without any kind of testing for yourself?

It is absolutely possible to produce FULLY realistic renders in a matter of a few minutes, you just have to have the scene correctly lighted, with a good GPU and some nodes setups. Viola.

That 9 hours was for a particular scene, with particular lighting setups that are not “average”. I have seen the image, the way the lighting is set up there would have been a lot of noise (very small lights, causing a very difficult situation for the algorithms) early on, and so would be baking for a VERY long time.
However, a scene with big lights, an environment map (especially with Farny’s environment map light sampling changes/fixes) will make a render nearly noiseless in a few minutes. All up to what the scene requires!

Pulling numbers from somewhere else has no real bearing on anything. It is the same thing with BI, if a scene has 10 million + polys, with complex lighting, AO, ray tracing, etc. etc. etc. It is going to take FOREVER to finish, why should it be any different with Cycles? Or any other rendering engine for that matter?

If you make a material and lighting setup that is as complex as you would do in BI, you will get a similar speed.
Just want to point out one thing, Once a BI render is done, it will be noiseless(Unless you set very few samples for env light/AO).
With Cycles, it’s a neverending render, and the more you render, the slower it gets. (i.e. passes 10-20 will do less help than passes 1-10). I’ve done several animation projects with it, and I’m never happy with the render result(I’m GPU challenged, and I can only afford about 150-400 passes/frame at max).

Did you try turning off GI bounces, and using only lights with size of 0?
Yes, the shadows become sharp, but what BI does is simply blur the sharp shadows instead of accurately raytracing them. This could probably be achieved in the compositor.

blur the sharp shadows instead of accurately raytracing them. This could probably be achieved in the compositor.
I can think of many things wrong with doing that after in the compositor. Lack of shrinking Umbra, shadow bleeding…

One thing most people simply don’t do in BI is glossy reflections, even though every object that has specular highlights should technically have (glossy) reflectivity on it too. Glossy reflections in BI take forever to render without noise and you will have to spend a lot of time figuring out the right setting that is both noise-free and doesn’t make your render time explode completely. Alas, cycles being physically based does not leave you the option to not do these reflections.

With Cycles, it’s a neverending render, and the more you render, the slower it gets. (i.e. passes 10-20 will do less help than passes 1-10).
It doesn’t get slower, but it is indeed diminishing returns. It’s as simple as this: Four times the render time equals half the error rate (noise).

I can think of many things wrong with doing that after in the compositor. Lack of shrinking Umbra, shadow bleeding…
Which again brings us to the point that shadow (maps) in BI were fast-but-generally-not-fantastic anyway. I for one am very excited by the tradeoff in light and shadow quality to time. Still very keen in experimenting with Cycles for animation purposes, speedups, what node setups are noisier than others, etc.

Once Mango is done, I’d love to see an Open Movie Training disk or similar to do with Cycles production style hints. If that doesn’t come about, collecting all the random hints that are already on the forum and condensing them into the Blender Wiki somewhere would be handy.

A bit hesitant to spearhead such a thing now though, given that in the next 6-12 months, Cycles will probably change around a lot with Mango. Excited to watch the progress, especially with the 4k requirements… :slight_smile: Worth noting that they might only need to render ‘an element’ of a 4k frame that isn’t anywhere near 4k res itself for many shots. Still, a challenge that I’m sure Brecht and co will rise to.

I know what you mean. I just finished my first attempt at a full scene Cycles rendering. The scene wasn’t that involved and it took roughly eight hours to render and that was at 1000 samples. Even if I had a newer, faster computer, which I do not, the rendering time would be prohibitive. Rending an animation of any length is out of the question without having access to a render farm. Rendering the same scene in BI would have taken an hour to render at the most. However, the final product rendered in Cycles was impressive to say the least. The light/shadow effects were so detailed, it’s almost impossible to distinguish them from the real thing.
I will continue to use it where I can. I don’t think it’s appropriate for everybody or every project situation. If you don’t have a shiny new super computer or your own personal stack of servers for rendering, using it regularly is going to be difficult or impossible for most. However, it’s nice to have it there when you need it…:wink:

My problem isn’t the time it takes to render Cycles but the difficulty in converting old Blender files to the new nodes setup.

It means retexturing everything all over again which is very time consuming and it’s difficult to get the same texture setup for some scenes, but I will persist because overall I think Cycles is much better. Just painful having to learn how to set up everything using nodes.

Is that sentence correct? I mean, is it a sort of mathematical rule or something like that going on in Cycles? Quadratic-Inverse time vs noise?

Actually, with normal random generator formula is 1 / SQRT(N) N = number of samples. It is well known formula from Monte Carlo integrator theory.

But with low discrepancy quasi random generator (already used by Cycles, it based on Sobol sequence) that curve become more better and close to linear.

BTW, i have nice progress in bi-directional part, at last such scenes now converging much faster, orders of magnitude (pure light to camera lights chains used). It will help a lot to clean noise near strong light sources, in participating media as on that image or just lamp close to wall.

That image take 40 minutes because i forgot to stop it, in reality it look nice after 20 samples, 1 minute is enough for animation, with some despeckle 2D filter of course.

Left black horizontal glowing plaine actually emission quad, i have not done yet light contribution directly from emitters to camera.

Indeed there is!