Why the polygon diminishing returns picture is bad

The Omega Man

Member
Jan 19, 2010
8,312
0
585
Montreal - Canada
It is pretty known that polygons count is not everything, let alone it should never be a measure of a great model or graphics. But video game programmers know this already.
 

SPE

Member
Mar 30, 2007
6,607
1
0
We're a long, log way from diminishing returns in gaming being a thing.

Just look at CGI in the movies. The same diminishing returns can be said about that, but GCI ages badly. State of the art effects on older films, like in the LOTR trilogy, can now look pretty ropey. Real advances are made year on year in film CGI. And consistent photo-realism is still just out of grasp.
 

sangreal

Member
May 19, 2005
36,226
0
1,225
34
but that still shows diminishing returns, without the conspicuously missing 40k polygon model. diminishing returns doesn't mean no room for improvement
 

RaikuHebi

Banned
Jun 16, 2013
8,576
0
0
This takes me back to my childhood. Talking about which console had more polygons. Good times.
 

nakedeyes

Member
Nov 10, 2013
660
0
335
Salt Lake City, UT
Obviously, more polygons authored by HAND will help with detail..

Maybe I'm missing part of the picture, but is the subdivision stuff simply talking about the new focus on tessalation as of late?

Cause that tessalation isn't primarily to add detail to the Geometry, it's purpose is to help with the lighting. As most all lighting algorithms not on pixel shaders, give much better results with many small polys, than one big one.

Tessalation is all about lighting as far as I'm concerned, and occasionally helps ( or apparently detracts ) with the geom's look.

-confused game programmer
 

DICKS AHOY

Banned
Nov 4, 2010
9,638
0
0
That doesn't really negate the point that after a certain point the increased polygon count becomes more negligible at a casual glance. Especially if in the first pic the 2K model is the original, given the 4K one next to it still doesn't look significantly better.

Think it goes without saying you won't magically wind up with a bust of Beethoven if you keep subdividing a cube in Maya/3DS Max/whatever.
 

XANDER CAGE

Member
Mar 21, 2011
21,046
0
0
The reason the 20k and 40k versions look so similar is because the base model itself looks to be incredibly simple. Look at how much detail the collar and buttons lack. The collar itself doesn't even actually fold, it's essentially just a collar-shaped divot in his coat.

In the future (hell, now, in non-realtime rendered things like CGI movies and art renders) you could probably put 5k tris into a really nicely modeled coat button.
 

Soupstorm

Neo Member
Jun 26, 2013
205
0
0
That doesn't really negate the point that after a certain point the increased polygon count becomes more negligible at a casual glance. Especially if in the first pic the 2K model is the original, given the 4K one next to it still doesn't look significantly better.

Think it goes without saying you won't magically wind up with a bust of Beethoven if you keep subdividing a cube in Maya/3DS Max/whatever.
It's still kind of useless because we still see clear polygonisation in practically every game. Yes, there is a theoretical limit to the aesthetic utility of more polys, but we are not anywhere close yet, and it's only going to get worse as display resolution standards increase alongside expanding game worlds.
 

Muffdraul

Member
Apr 23, 2008
10,199
16
915
Los Angeles
but that still shows diminishing returns, without the conspicuously missing 40k polygon model.
Thank you, was about to point this out. Very convenient that they left out the new 40k model. The 20k to 40k comparison is basically the punchline that makes the point.
 

EatChildren

Currently polling second in Australia's federal election (first in the Gold Coast), this feral may one day be your Bogan King.
Jan 29, 2008
36,153
0
0
Australia
While I don't agree with the "diminishing returns" argument due to other variables in 3D interactive rendering that are routinely ignored in analysis, this refute is also pretty poor. Much of the detail added in their version of the 20,000 poly model is already replicated with advanced techniques like normal mapping, giving the illusion of real 3D geometry interacting with light/shadow where there isn't actually any. A better example (though I can appreciate the difficulty of constructing this) would be densely packed full scene of geometric detail, as however many triangles you can push in a single frame render isn't limited to individual assets but instead the entire scene. Having a higher poly budget lets you add greater geometric detail to any one scene.
 

Neat Machine

Member
Jun 2, 2007
8,579
31
915
The point of the picture is to show that the "diminishing returns" claimed by the original image is mostly an effect created by where the source image is located.

It's illustrated pretty clearly when the source is changed to the first image. Overall, what you're seeing is that arbitrarily throwing polygons at something doesn't make it look more detailed. You have to actually add the detail, which a higher poly count allows you to (but doesn't automagically) do.

Edit: I agree that the point of the original picture was flawed from the beginning, but it's funny to see that even the concept of it was executed so badly.
 

rrs

Member
Nov 18, 2012
6,459
0
0
Not Chicago, IL
While I don't agree with the "diminishing returns" argument due to other variables in 3D interactive rendering that are routinely ignored in analysis, this refute is also pretty poor. Much of the detail added in their version of the 20,000 poly model is already replicated with advanced techniques like normal mapping, giving the illusion of real 3D geometry interacting with light/shadow where there isn't actually any. A better example (though I can appreciate the difficulty of constructing this) would be densely packed full scene of geometric detail, as however many triangles you can push in a single frame render isn't limited to individual assets but instead the entire scene. Having a higher poly budget lets you add greater geometric detail to any one scene.
I was about to say that, polygons don't mean as much as they used to - how you use them is key. A dynamic physics bound trenchcoat will look better that one glued to a model's legs.

Nostalgia! Loved Interstate 76!
I like how the devs tried to make the prerendered sequences look like could be done on the consumer hardware of the time, it still looks good while the era's "totally realz" 3D cousins fall flat.
 
Sep 11, 2007
26,921
0
0
Thank you, was about to point this out. Very convenient that they left out the new 40k model.
I think the whole point is that the original 40k mesh used in the example is a bad example, since most of the polygons are "wasted" they don't add more detail like it could do. It just increasing the polycount form a basic model.

As it shows later, you can add a lot of detail with more polygons that aren't exactly reflected on the 2k mesh. So more polygons, do actually matter.

Edit: Although like EatChildren said most games uses normal maps, but I think is just a valid refute of the Polygon argument
 

beril

Member
Aug 25, 2010
3,993
0
0
Yes the original image is flawed but proves a point that this really doesn't do a good job of disproving.

Also if you're not using flat shaded untextured polygons it becomes even more true. All the extra detail they added in the 20000k version could very easily be done in a normal map without much noticeable difference. The same is not quite true between the 200 and 2000 triangle version
 

SneakyStephan

Banned
Jan 23, 2011
18,402
0
0
As long as games have LOD pop in and visible lod transistions for geometry and plates/wheels that look like decagons up close and other illusion breaking shit we are nowhere near hitting any diminishing returns for polygons on screen.

Maybe in another 20 years
 
Apr 27, 2009
9,969
1
0
Bay Area
OP ever hear of


You are subdividing a poor mesh with really poor topology. If you want to SUBDIVIDE, you use quad modelling or sub-d modelling workflow (another way of quad modelling).

Zbrush actually has a pretty decent poly reducer and Mudbox is also amazing.

If there is any takeaway I can say about the next gen is that polygons are king. You have to bevel everything.
 

EatinOlives

Member
Oct 23, 2011
16,850
0
0
That doesn't really negate the point that after a certain point the increased polygon count becomes more negligible at a casual glance. Especially if in the first pic the 2K model is the original, given the 4K one next to it still doesn't look significantly better.

Think it goes without saying you won't magically wind up with a bust of Beethoven if you keep subdividing a cube in Maya/3DS Max/whatever.
But that's what the original picture was trying to do. By subdividing the picture you can claim you're "increasing polygon count" even though you're not adding any real detail that increasing the polygon count actually lets you do.

"After a certain point" is the whole debate. What's this "certain point"? There are people who like to claim the graphics output of their current console of choice is that "certain point". It's silly, though, because 5-6 years down the line it's very obvious we're nowhere near that point. Those who started throwing the diminishing-returns argument were doing so as early as the PS2 days, which in hindsight is extremely silly.

And it doesn't take too much to realize we're nowhere hear that "certain point". One look at any Blu-ray Dreamworks/Pixar movie of your choice from the last ~4 years will tell you that much.
 

sirap

Member
Mar 1, 2013
6,109
0
0
Malaysia
it's far more efficient to represent surface detail with normal maps and focus your polygon budget on getting clean sillhouetes and model topology.
 

Muffdraul

Member
Apr 23, 2008
10,199
16
915
Los Angeles
I think the whole point is that the original 40k mesh used in the example is a bad example, since most of the polygons are "wasted" they don't add more detail like it could do. It just increasing the polycount form a basic model.

As it shows later, you can add a lot of detail with more polygons that aren't exactly reflected on the 2k mesh. So more polygons, do actually matter.

Edit: Although like EatChildren said most games uses normal maps, but I think is just a valid refute of the Polygon argument
I think their point would be much stronger if they did all four models, same as the original. But really, I think that would have destroyed their point, because the 20k and 40k would surely have been just as hard to distinguish. If not, why not show them? "We ran out of room on the jpeg!" Hmph.

People take the original far too "literally". It's an abstract example that illustrates what diminishing returns means, and explain why it's only natural that we're not going to be as blown away going forward as we were in the past. Sure, it can be mitigated by whatever new fancy graphics tech we have now on top of the polygon count, but we didn't need that to see an obvious difference between Star Fox, Colony Wars, and Rouge Squadron 2.
 

CyclopsRock

Member
Jul 6, 2012
12,297
0
490
In CGI, one of the problems with high poly counts isn't GPU based but RAM based, but even that's far less a problem than it used to be. what IS still s problem, though, is mesh deformation. it's very common for a TD to give a character rig to an animator with a super basic model attached, and then swap it out for the final model just before render time. the reason for this is that the animator wants a nice, quick response to their changes and viewport, and high detail mesh still tends to deform slowly. This, I think, especially when you need it real time, might be a reason why you still see the polygons on characters - because the GPU being able to render it isn't everything.
 

Alchemy

Member
Sep 1, 2008
8,638
0
0
There are still diminishing returns, it just isn't as bad as the original picture shows. The gap closes significantly when normal mapping is applied however, and texture resolution has a huge impact.
 

RamsesA

Neo Member
Mar 10, 2013
119
0
0
I never liked the infographic because it completely ignores normal mapping, which is a common way to add bumps and details to an object without using more polygons.
 

Orayn

Member
Jul 4, 2010
34,290
3
690
Minnesota
Efficient yes, but it still looks(alot) worse than real geometry.
The next big mountain we need to climb is parallax occlusion mapping. It's kind of like tesselation, but with a much, much lower performance cost because the added detail is "fake."

 

Mr_Appleby

Member
Mar 21, 2008
3,940
0
765
While I don't agree with the "diminishing returns" argument due to other variables in 3D interactive rendering that are routinely ignored in analysis, this refute is also pretty poor. Much of the detail added in their version of the 20,000 poly model is already replicated with advanced techniques like normal mapping, giving the illusion of real 3D geometry interacting with light/shadow where there isn't actually any. A better example (though I can appreciate the difficulty of constructing this) would be densely packed full scene of geometric detail, as however many triangles you can push in a single frame render isn't limited to individual assets but instead the entire scene. Having a higher poly budget lets you add greater geometric detail to any one scene.
normal mapping is still only a partial solution. It simulates the effect of light on a surface that is different to the actual polygonal surface, however when viewed from certain angles it becomes readily apparent that you're looking at a flat surface with a special shading, as opposed to a genuine detailed surface. The recent shoulder images from the NBA game shows off this effect an apparently rounded surface from some angles that when viewed from others is revealed to be far less smooth. I'd post a better example if I wasn't on my mobile - it can easily be seen if you compare a sphere normal mapped to look bumpy with one that has bumped geometry. The bumps in the normal mapped one will not protrude beyond the sphere edges the same way 'real' bumps will.
 

Speevy

Banned
Jun 26, 2004
62,861
0
0
Am I the only person who read the words "polygon" and "bad" together, and immediately thought of the site?
 

Ishan

Junior Member
Jun 13, 2013
5,048
0
0
Am I the only person who read the words "polygon" and "bad" together, and immediately thought of the site?
nope. I thought this thread would probably be an article from polygon about diminishing returns.
 

Margalis

Banned
Aug 7, 2008
2,954
0
0
This picture doesn't really illustrate anything other than that programmatic tessellation of a 200 polygon model is a dumb idea. In fact it does illustrate diminishing returns, as in the gap from 20k vs 2k polygons is very low in visual fidelity compared to 200 vs 2k.
 

Narroo

Member
Jun 6, 2013
2,073
33
350
The next big mountain we need to climb is parallax occlusion mapping. It's kind of like tesselation, but with a much, much lower performance cost because the added detail is "fake."
I...don't see much of a difference.


But more importantly, what people are confusing is disproving a proof with disproving a theory. I could claim the sky is blue "due to the sky reflecting the color of the ocean," and be horribly wrong, but the sky is still blue.

The original post is great because it destroys a poorly written example, but it doesn't destroy the theory. Rather, it simply removes the original example from the argument.
 

forrest

formerly nacire
Dec 9, 2008
5,109
0
0
Seattle, WA
forrestcrump.wordpress.com
I don't know that the example shown really portrays diminishing returns of polycount. However, it is a great example of why cross platform assets should be created for the highest fidelity platform and then optimized down.

*Additionally, I've never specifically looked into how tessellation is calculated in modern gaming, but I would imagine that this would also stand as a good example of why extremely high levels of tessellation would be a bad thing.