Ok, thank you
You, like lukas stockner getting hung up on the trivialities of the scrambling distance patch represent a non-creative take on what is (often) a fundamentally creative technology. Its important but only so much. The ability to have fast and satisfying proxies/samples or even slightly (or very) stylized output frames is hugely significant. I think you know this but you are hung up on projecting ignorance onto others. Your points are well made but I have to ask - are you an artist? You sound like a pathologist talking about the clarity of your scans not an artist with vision for their own creative contributions. I can see you 150 years ago trying to tear down impressionism because its not realistic, lol.
Yes, because like Lukas, I also know what I am talking about. I recognize the importance of being able to spend most of the time focusing on art rather than tweaking dozens of technical parameters which balance speed and bias in all the various different ways. Artist time is way more expensive than computer time.
I am not saying what bliblubli is doing is wrong. Go take look at my first post of this argument. All I asked about is that I’d like to see how his optimizations perform in a setup that doesn’t have lighting faked to such a degree it looks like a frame from a game. I just wanted to see if the speed difference applies also for use cases where user can’t afford to compromise quality to gain speed.
And yes, I’d consider myself an artist: https://www.artstation.com/rawalanche
Here are some more benchs i did using new intel denoiser:
-Tiles Size 16
Always Rendering with:
-GPU: 2x GTX1080 (cuda)
-CPU: I9-7940 (using 28 Threads)
1000 Samples/ mblur
1000 Samples/ mblur
800 Samples/ mblur/ Denoise
800 Samples/ mblur/ Denoise (intel/comp)
200 Samples/ mblur/ Denoise
200 Samples/ mblur/ Denoise (intel/comp)
…but the general consensus is that Lukas is wrong in his take. He is a perfect example of tunnel vision that ignores even the technical director on some of the most prominent uses of Blender to date. And thats in a traditional realism based useage - not even getting into more fine art / stylized uses of Blender.
And I see that you are now moving the goalposts/diminishing how aggressively you started because you see you overdid it. Thats good.
But this idea that a render that looks painted at worst:
A. resembles an EEVEE render in any way (artifacts are entirely different)
B. is a direct indicator of “low quality” to anyone, anywhere
C. isn’t eminently helpful while developing work
are all completely fallacious.
All that being said - you have very nice work!
A, No they aren’t. Eevee isn’t capable of producing accurate enough global illumination, so it relies on ambient occlusion for contact shadows. That’s exactly what I am seeing on that Cycles render.
B, Yes, by today’s standards of lighting quality it’s considered low quality.
I am talking about the lighting quality, not the other aspects, such as modeling or texturing work. They are actually on good level, which makes the lacking lighting quality that much more apparent.
C, No, the final workflow isn’t much helpful, if you have to switch 10 different knobs every time you jump between preview quality and final quality. All the more problems arise when you are not working in the conditions of final image. We should be pushing towards interactivity and WYSIWYG workflows, not away from them. Having different settings for previews and finals is legacy workflow that should die as soon as possible.
On top of that, there ARE solutions which can make Cycles several times faster, and at the same time actually increase the resulting quality of the artwork by allowing artists to use more realistic amount of indirect light bounces without worrying about destroying performance. More on that here: Cycles Performance
But I still strongly believe that achieving that by regressing back to ancient primitive methods like constant ambient light with small, strong ambient occlusion shadows is not a way to go about it. It generally causes users to produce lower quality work while taking even more time to do so.
Well, I’m not sure it’s what you want to say, but anyway, E-Cycles doesn’t use any AO trick to be faster. It’s an option because as you can see, different artists have different needs and different clients have different budget. So it’s also not about having preview quality and final render quality with 10 knobs to move: If you like photo realism, you set 128 bounces everywhere, disable clamping, disable filter glossy, etc. and E-Cycles will render that 2x faster than master.
Yes, absolutely. I am not arguing about that. What I mean is that based on that picture, Evermotion guys, or whoever has set up that scene, decided to use AO. All I wanted was just to see how the speed difference would look in a scene where such fakes as AO/simplify bounces are not used
I’ll give an example soon
I’m ok with you all discussing what is good art in the E-Cycles thread, but what you say give the impression to many that I use biased tricks in E-Cycles by default. It’s not the case. The 2x faster render gives the same results as master, whatever the settings are. Using tricks, it becomes 4 up to 10x faster and I showcase that also because some are going to Eevee and my point is to show them Cycles can be as fast with better results.
And on top of that, it’s much easier to switch between cheap rendering to HQ rendering in one engine then to be stuck in an engine that can only do cheap renders
Yes, again, not disputing that.
I’d just like to see some comparison. For example Cycles in master vs E-Cycles in scene with no fakes. And then E-Cycles with no tricks vs E-Cycles with tricks. With both rendertimes and side by side pictures. That will make it a lot easier to judge what kind of speed/quality tradeoff is actually going on.
The whole reason I started this conversation was just to let people know that those bombastic 20 second rendertimes do come at a price. Even though it may be not that obvious on that particular image to everyone, it may come to bite them in the back in some other scenes, which rely on accurate light transport a bit more.
Renders are coming
Thanks for the test. You have a very fast CPU, but E-Cycles speeds your GPUs, the CPU still render as fast as before. So it makes the overall speedup smaller. On top of that, the new embree code for faster motion blur is now in master, but only for CPU as far as I know.
I’m working on a speedup for CPU, so it should become better soon.
This is the original version of the scene number 7 from this evermotion pack. It renders in 18min10sec on master and 13min07 on E-Cycles (pixel perfect, exact same render):
- a limitation of 4 bounces for diffuse and glossy
- is limited to 12 bounces total (transmission+diffuse+glossy and transparency)
- has none of the 2 caustics
- has both clamp at 10.
So now here 2 extreme possibilities of E-Cycles:
Use the faster rendering to keep about the same render time, while dramatically increasing photorealisme:
- all GI bounces at 16 (vs 4 before)
- full dynamic range (vs both direct and indirect clamping at 10 before)
- both reflective and refractive caustics on (vs both disabled before)
- it uses the new sampling algorithm of E-Cycles, but the noise is kept at the same level by reducing samples.
-> 9min56 compared to 18min10 = 1,82x faster, but requiring changing the spp value and activating the new sampling (2 clicks, one keyboard input). If you just want the lazy version, with same spp and same sampling as original, it’s as fast as master with the original.
Use E-Cycles to render very fast:
All the same settings as the improved quality one + AO simplify at 4
-> 2min32 compared to 18min10 = 7,28x faster. Note how it is actually much nearer to the full GI render than the original file using master.
Of course, with AO at 2 and using more tricks, you would still get something superior to Eevee and also around the same render time.
Your advertising tagline is
Render Interiors in Under a Minute With CUDA.
People who read that will expect that if they pay for the fork, they can render similar results to what is possible in Vray, Corona, and Arnold in under a minute without doing too much. There is no ‘*’ symbol or anything that indicates it coming with the catch that tricks need to be activated.
I’ve been sitting on this question for a while, what is the plan here to keep people paying from month to month? I know they get a patched Blender, but they also have the code and there’s nothing under the GPL that prevents them from submitting a tweaked version to the BF, or from providing their own builds for free.
& well, to my eye the second one (faster) looks lousy - cheap & flat, from archviz PoV. And for animations time is still unacceptable.
As i already mentioned to you, when you were promoting fast CUDA/OCL builds ~ year & a half ago - i get better & faster results with cycles of my own tricks. But now you’re even selling “a cat in a sack” for quite a sum… . Also, there’s no “perpetual license”, only GPL for Blender - correct your gumroad page!
Two and half minutes is unacceptable? Seriously?
Well, you got me totally biased exaggeration … since i don’t have a scene to test, nor did i look for full specs - only checked the resolution and time, i assumed from my personal experience.
If anyone dislikes it or finds it offensive - flag it!
@bliblubli: I think you’re doing a great job. Buyers can find all the info they need inside this topic, if someone is not able to read and understand data that’s another problem. Same is if someone blindly buy a product after reading a title, when there’s so much info about it. While some of the discussions here are constructive, I see also a lot of posts with a lot of unnecessary aggression and blind talk. I hope you won’t get too influenced by that and keep up the great work you’re doing.
Keep in mind that the ability to make pathtraced animation on home hardware (with all of the accuracy) will be a huge testament to the advances in computer hardware and rendering technology. It’s already bordering on achievable with the GTX 2xxx series, the Vega VII, and the AMD threadripper WX.
It’s not uncommon for the big Hollywood blockbusters to have VFX that takes many hours for one frame (but they have huge server farms that can render hundreds or even thousands of frames concurrently).