Search - User list
Full Version: arnold or redshift? h18
Root » Houdini Lounge » arnold or redshift? h18
anon_user_37409885
@Darly - nice to see you are dropping your claim then:

Daryl Dunlap
GPU is the future for Production render engines
Midphase
I ended up buying Redshift after watching this video:




But in all honesty I've been kinda having Octane/Arnold render envy, especially after seeing some of the more impressive examples of what they can do.

Ultimately, once RS becomes Metal-compliant I'll go back to it since for some stuff the speed considerations override the accuracy, and it is pretty mature in terms of nodes etc.

I'm currently using 3Delight (if it wasn't clear by now!), and I really dig the speed, but it's got ways to go in terms of matching the wealth of nodes that most other renders offer.
stephan6
ok, i think i have to chip in here i've used Arnold in production since 2009 and Redshift pretty much since v1.0. Both in Softimage. First of all, Arnold felt like a revelation back then coming from MentalRay (even on old CPUs were it was veeeery slow). Redshift had pretty much the same effect when it surfaced, I was finally able to get a shot rendered on a single day. (yay )

Both use pretty much the same algorithm and shortcuts, so i doubt that you'll see a big quality difference there. One thing i noticed, is that RS renders bumpmaps differently (worse), especially procedural textures. Arnolds bumps are sharper, less oversampled, but still nicely antialiased. in RS you have to drive the reflection roughness instead to get a similar look, so it's doable. But you'll have to optimize more, the days of simply adding a Arnold property with 5 additional subdivs for displacement are over

there are a lot of excellent RS projects out there, just check out panoply's work. (https://www.panoply.co.uk)

if you see a difference in this fence/sparks scene, it's probably just because RS clamps secondary rays (which is optional), to avoid fireflies and overcast pixels. I didn't find a similar setting in Mantra, and had therefor pretty badly sampled renderings with it. (but i'm not very experienced in Mantra) (and lets not talk about the speed, especially with transparencies)

Either way, I try to get back to CPU, but mostly because i want to use bidirectional pathtracing, and i'm sick of the way windows behaves while rendering on GPU (especially when it goes out of core, my system is barely useable - so i can't really render in the background… which isn't the case with CPU renderers). The speed improvement of RS also diminishes when it's getting very complex, and the scene extraction times of RS aren't that good either to tell you the truth.

I'm having a hard look at 3delight at the moment.. and it's a neat experience so far…. there are still a lot of features missing (it's in its early days), but the speed is almost on RS level, and the plugin (even in beta) is rocksolid compared to Renderman. (not bidirectional though)

just my 0.02$

–Stephan Haidacher
anon_user_89151269
if enough time passes, every rendering engine under the sun will be be mentioned here as a preference.
but I'm stuck on that stark difference between the RS rendered sequence and the Mantra one, in that example posted by Daryl. Am I the the only one seeing the difference?
Midphase
No you're not. Mantra looks far superior as far as realism.
TwinSnakes007
RS, by default MipMaps textures, and uses a texture size based on pixel coverage to save VRAM space, but you have full control over how aggressive the MipMapping behaves. So, for Hero shots, you can fully disable MipMapping to achieve better texture detail, for example, BumpMapping.
TwinSnakes007
Midphase
No you're not. Mantra looks far superior as far as realism.

…and Maxwell would look even more superior…but, nothing is free (in terms of frametime).

You could render that animation in every engine that exist, and on their own, I'm sure they would look acceptable - the major difference would be the frametime.
stephan6
Daryl Dunlap
RS, by default MipMaps textures, and uses a texture size based on pixel coverage to save VRAM space, but you have full control over how aggressive the MipMapping behaves. So, for Hero shots, you can fully disable MipMapping to achieve better texture detail, for example, BumpMapping.

i'm very aware of the mipmap bias setting, but this doesn't solve the issues i have with RS bump mapping. especially with procedurals

–s
anon_user_37409885
stephan6
there are a lot of excellent RS projects out there, just check out panoply's work. (https://www.panoply.co.uk)


Thanks, this looks great but it confirms the ‘plastic’ look of RS shaders. I've never seen a job from RS that doesn't look dullish.
intoArt
stephan6
First of all, Arnold felt like a revelation back then coming from MentalRay (even on old CPUs were it was veeeery slow). Redshift had pretty much the same effect when it surfaced, I was finally able to get a shot rendered on a single day. (yay )

Both use pretty much the same algorithm and shortcuts, so i doubt that you'll see a big quality difference there.

Sorry for being a bit harsh, but your introduction sounds like Redshift and Arnold produce similar results and Redshift is just faster which is pretty much misleading.

stephan6
I'm having a hard look at 3delight at the moment.. and it's a neat experience so far…. there are still a lot of features missing (it's in its early days), but the speed is almost on RS level, and the plugin (even in beta) is rocksolid compared to Renderman. (not bidirectional though)

I also don't understand all the hype about render engines. Compared to your statement above 3delight doesn't use the same algorithms than Arnold / Redshift and does something amazingly better ? and Renderman is not good enough in your opinion ? Thats a bit arbitrary. Arnold and Renderman are amazing and both get gpu or even better xpu support in the future.
stephan6
intoArt
stephan6
First of all, Arnold felt like a revelation back then coming from MentalRay (even on old CPUs were it was veeeery slow). Redshift had pretty much the same effect when it surfaced, I was finally able to get a shot rendered on a single day. (yay )

Both use pretty much the same algorithm and shortcuts, so i doubt that you'll see a big quality difference there.

Sorry for being a bit harsh, but your introduction sounds like Redshift and Arnold produce similar results and Redshift is just faster which is pretty much misleading.

stephan6
I'm having a hard look at 3delight at the moment.. and it's a neat experience so far…. there are still a lot of features missing (it's in its early days), but the speed is almost on RS level, and the plugin (even in beta) is rocksolid compared to Renderman. (not bidirectional though)

I also don't understand all the hype about render engines. Compared to your statement above 3delight doesn't use the same algorithms than Arnold / Redshift and does something amazingly better ? and Renderman is not good enough in your opinion ? Thats a bit arbitrary. Arnold and Renderman are amazing and both get gpu or even better xpu support in the future.

misleading? so tell me, what's the difference between both pathtracers? (keep in mind, when RS was released Arnold didn't have randomwalk SSS, or the volume support it has now, which is the main feature difference atm.)

and yes Renderman 22.6 and 23.2 are buggy as hell in Houdini on Windows. I worked with it for around 4 months and had more crashes than in my 25 years combined. It was one of the most frustrating experiences ever, so I wouldn't call my opinion arbitrary. anyway, use whatever works for you.
anon_user_37409885
stephan6
misleading? so tell me, what's the difference between both pathtracers?

It's not opensource, so you don't know what the code is doing.
stephan6
lol, you guys should read about montecarlo based pathtracers….. but i should have known better to argue with experts on a forum. i'm outa here
anon_user_37409885
stephan6
lol, you guys should read about montecarlo based pathtracers….. but i should have known better to argue with experts on a forum. i'm outa here

What do you recommend to read?
mandrake0
goat
stephan6
lol, you guys should read about montecarlo based pathtracers….. but i should have known better to argue with experts on a forum. i'm outa here

What do you recommend to read?

Technial:
https://www.pbrt.org/index.html [www.pbrt.org]

A bit offtopic but also a intestting read. JIT based render engine. https://rgl.epfl.ch/publications/NimierDavidVicini2019Mitsuba2 [rgl.epfl.ch]

There are a lot of ressources from shader to algorythms….
anon_user_37409885
Thanks but that's not the implementation. Each renderer does shortcuts and optimisation, otherwise, they would be carbon copies of each other, in imagery, speed etc.
Midphase
intoArt
I also don't understand all the hype about render engines. Compared to your statement above 3delight doesn't use the same algorithms than Arnold / Redshift and does something amazingly better ? and Renderman is not good enough in your opinion ? Thats a bit arbitrary. Arnold and Renderman are amazing and both get gpu or even better xpu support in the future.

FWIW:

To me render engines are just like video cameras. The sensors are all based for the most part on the same technology, yet they yield vastly different results (when I'm not working on my personal Houdini projects, I do color grading for clients). For instance, the way an ARRI Alexa shot project looks and “feels” vs. one that was shot on a Sony FS7 is night and day. In addition, when operating a camera, the way the menus are laid out, what knobs and buttons and placed where, etc. makes a huge difference for the DP's and operators.

Render engines also each have a look and feel to the images that they produce, and the equivalent of the menus/knobs and buttons are the type of nodes that they offer and how they are wired between each other. In addition to the above, there is also a time-to-render factor, which for small one-person-armies can be a huge consideration.

I think this is exactly why the hype about render engines.
jarenas
Midphase
intoArt
I also don't understand all the hype about render engines. Compared to your statement above 3delight doesn't use the same algorithms than Arnold / Redshift and does something amazingly better ? and Renderman is not good enough in your opinion ? Thats a bit arbitrary. Arnold and Renderman are amazing and both get gpu or even better xpu support in the future.

FWIW:

To me render engines are just like video cameras. The sensors are all based for the most part on the same technology, yet they yield vastly different results (when I'm not working on my personal Houdini projects, I do color grading for clients). For instance, the way an ARRI Alexa shot project looks and “feels” vs. one that was shot on a Sony FS7 is night and day. In addition, when operating a camera, the way the menus are laid out, what knobs and buttons and placed where, etc. makes a huge difference for the DP's and operators.

Render engines also each have a look and feel to the images that they produce, and the equivalent of the menus/knobs and buttons are the type of nodes that they offer and how they are wired between each other. In addition to the above, there is also a time-to-render factor, which for small one-person-armies can be a huge consideration.

I think this is exactly why the hype about render engines.
I could not have explained it better!
stephan6
Midphase
intoArt
I also don't understand all the hype about render engines. Compared to your statement above 3delight doesn't use the same algorithms than Arnold / Redshift and does something amazingly better ? and Renderman is not good enough in your opinion ? Thats a bit arbitrary. Arnold and Renderman are amazing and both get gpu or even better xpu support in the future.

FWIW:

To me render engines are just like video cameras. The sensors are all based for the most part on the same technology, yet they yield vastly different results (when I'm not working on my personal Houdini projects, I do color grading for clients). For instance, the way an ARRI Alexa shot project looks and “feels” vs. one that was shot on a Sony FS7 is night and day. In addition, when operating a camera, the way the menus are laid out, what knobs and buttons and placed where, etc. makes a huge difference for the DP's and operators.

Render engines also each have a look and feel to the images that they produce, and the equivalent of the menus/knobs and buttons are the type of nodes that they offer and how they are wired between each other. In addition to the above, there is also a time-to-render factor, which for small one-person-armies can be a huge consideration.

I think this is exactly why the hype about render engines.

well, that's a romantic thought

the simple problem with your comparison is, we are not talking about sensors and lenses capturing light (analog), we are simply talking about math. renderers work linearly, use the same rendering base algorithm (http://graphics.stanford.edu/courses/cs348b-10/lectures/path/path.pdf), use mostly the same shading models (GGX), use the same pixel filters (gaussian for example) and use the same light falloffs (squared).

the big differences are feature sets, speed, efficiency and data handling (memory footprint). this is even true for bidirectional pathtracers.

look differences are mainly because of different shader feature sets, sometimes (but rarely today) texture handling(mipmaps), internal clampings, thresholds, applied LUTs or postFX (that's why octane renderings look nice from the beginning), but if you create a simple scene with the same object, the same light with the same intensity/size and same basic lightshader, the same basic surface shader with the same shading model, you'll get exactly the same result. (with an MC, rgb based pathtracer)

but of course renderers support different shaders with varying options, different light shader properties, and sometimes use different shortcuts to reduce the time to finish a frame. But i don't accept the premise of being “misleading” for comparing Arnold and Redshift.

a bit dated but still a good read
https://www.fxguide.com/fxfeatured/the-state-of-rendering/ [www.fxguide.com]
anon_user_37409885
@stephan6 thank you for rejoining the discussion. Unfortunately, it feels like you are are going down the same path as @Daryl.

A renderer is far more than the kernel of MCPT, it is the sum of all its parts, as you have pointed out.

As an analogy, you have been saying that running an Intel processor renders the platform of choice the same, because, the heart of the system is the same. MacOS, linux and Windows users would like to have a word with you.
This is a "lo-fi" version of our main content. To view the full version with more information, formatting and images, please click here.
Powered by DjangoBB