Hey all,

So although I've posted some stuff here for critique, by and large, I'm not really a CG artist who does stuff for games. To be fair, even CG art is primarily a side hobby for me. But anyhow, back on topic. One of the things I've noticed is that for general CG art, we generally prefer to use as much geometry as possible for building a scene, since we don't really care much about render time. For example, sometimes bricks will be modeled in to fully utilize the lighting calculations and what not. Now obviously, with game dev, that just isn't true, and I do have experience with this. What I've noticed, however, is that game art in many games (Dark Souls, etc.) really does seem to look like it's using a ton of geometry, especially with architecture. Now I can tell where textures are used, but I'm curious, what's the balance for geometry vs texture? When is a texture preferred to modeling geometry? 

I thought this would be interesting discussion. 

1 minute ago, deltaKshatriya said:

For example, sometimes bricks will be modeled in to fully utilize the lighting calculations and what not. Now obviously, with game dev, that just isn't true, and I do have experience with this. What I've noticed, however, is that game art in many games (Dark Souls, etc.) really does seem to look like it's using a ton of geometry, especially with architecture. Now I can tell where textures are used, but I'm curious, what's the balance for geometry vs texture? When is a texture preferred to modeling geometry? 

Often you do model everything in excessive detail, down to cracks in bricks, first. Then, second you make a lower detailed mesh with the same general shape (a flat wall that covers all the bricks). Third, you "bake" all the geometric details from mesh #1 onto normal/depth/ao textures for use on mesh #2.

If done right, mesh #2 now has all the same details as mesh #1, but hardly any polygons. You can only really tell the difference when looking at it extremely closely, or when looking at the silhouette.

As when when to use each... it can be easy to get carried away with using low-poly meshes with large textures, but textures take up a LOT of memory. A compressed 1024x1024 normal map is about 1.3MB -- in that same amount of memory you could store over 40K vertices! So it can actually be better to use extra polygons and lower resolution textures, where possible.

The main thing that you never want to do is end up with so many triangles that they start to become smaller than one pixel. GPU's like to run pixel shaders on a block of 2x2 pixels at a time, so if a triangle only covers 1 pixel, the GPU will likely run the PS on 4 pixels, throw away three results, and keep one result! This means that meshes that have an extremely high polygon count can actually increase pixel shading costs by 300%  Moreover, the fixed-function rasterizer probably works on 4x4 or 8x8 blocks of pixels, and only works at full efficiency if your triangles are bigger than that...

So -- use as many triangles as necessary, but then employ LOD (switching out models for lower detail versions when in the distance) to make sure that your on-screen triangles never get too small.

Maybe as an artist you should not distinguish between geometry and texture too much at all. There may be a preprocessing toolchain that takes your art and converts it to efficient game assets. It may resample textures, merge materials, decide what details remain geometry or become normal maps, generate LODs. It may even remesh your geometry and turn details to displacement maps, geometry images, voxels or whatever turns out to be most efficient.

So thinking ahead it's probably most important you always create art with very high detail so it still looks as intended even if everything gets resampled and downscaled. E.g. a tool may rotate UVs and straight lines of texels could become jaggy, or a tool may remove long and thin features of geometry, etc. Being aware of such issues may become important, while caring about low poly count may become obsolete for the artist.

I'm not a game artist by any means, I was just genuinely curious about the differences. It's interesting that both textures and geometry have limitations for game art. I knew about LOD tricks of course, since I've done some hobbyist game dev myself. 

I'd imagine that things like light fog, distance fog, camera angle, etc. can do a lot to change the visuals, potentially even hide some things? 

27 minutes ago, deltaKshatriya said:

I'd imagine that things like light fog, distance fog, camera angle, etc. can do a lot to change the visuals, potentially even hide some things? 

To me the perfect reference is looking SD or highly compressed video. It has only little detail and it's blurry, but it looks 1000 times better than any game graphics. It proofs we can remove a lot of things. ('We' means programmers - programmers need to solve this, artists should not need to care. Actually they spend way too much time on high to low poly conversion, creating UVs etc. That's why i believe in automated tools.)

This has been the problem with 3D graphics from the start, how many polygons can we use?

The answer is: no one knows.

There was a time where using alpha maps where used to make geometry look more complex; that was when computers had strict limits on polygons. These days computers can render millions of polygons but alpha sorting is much more expensive than it ever was; so doing the same thing will result with people calling you a fool.

So how do you know just how much geometry you can use?

Here is the rule of thumb: 64 000 vertices per draw call.(For optimized engines like Unreal and Unity).The thing is if you don't use the full 64 000 vertices per draw call it's still going to cost the same.

Except how many vertices each mesh has is hard to know. Because smooth groups and UV maps add to the account. So a simple object like a wall with bricks will have a simple almost flat UV map and no need for smooth groups, so can have lots more geometry. A motorcycle will need lots of uv maps for the complex shape and lots of smooth groups for the different materials, so will have less vertices.

Then there is the fact that there is no rule that one object should only have one draw call. Main characters in the game will often already have more than one material. In Dragon Age 2 the characters had many materials and so the characters had huge amount of polygons(> 182 000 even on low), the environment shared one material so objects in it like chairs where only in the hundreds(200 - 600 polygons).

It depends on the computer, the artist, the game type and the model. So no one really knows.

On 9/27/2017 at 7:38 PM, JoeJ said:

'We' means programmers - programmers need to solve this, artists should not need to care. Actually they spend way too much time on high to low poly conversion, creating UVs etc. That's why i believe in automated tools.

It's important for the artist to care, not caring would put a lot of extra pressure on the programmers and would allow the artist to shift blame. Team work is always important.

Auto tools can't do complex things yet and when they can they will be as smart as humans, so a artistic hand guiding it is important. Then there are the times where the auto tool makes such a mess you are better off without it.

On 9/27/2017 at 7:38 PM, JoeJ said:

To me the perfect reference is looking SD or highly compressed video. It has only little detail and it's blurry, but it looks 1000 times better than any game graphics. It proofs we can remove a lot of things.

Removing things is fine for videos where everything is batched, nothing is going to change suddenly.

So in a movie you can spend a few minutes compressing data, because you will only do it once. In a game textures have to be compressed at real time and uncompressed when used, so a heavy compression like the ones used in videos would reduce performance not gain it.

6 hours ago, Scouting Ninja said:

 

On 27.9.2017 at 7:38 PM, JoeJ said:

'We' means programmers - programmers need to solve this, artists should not need to care. Actually they spend way too much time on high to low poly conversion, creating UVs etc. That's why i believe in automated tools.

It's important for the artist to care, not caring would put a lot of extra pressure on the programmers and would allow the artist to shift blame. Team work is always important.

Auto tools can't do complex things yet and when they can they will be as smart as humans, so a artistic hand guiding it is important. Then there are the times where the auto tool makes such a mess you are better off without it.

On 27.9.2017 at 7:38 PM, JoeJ said:

To me the perfect reference is looking SD or highly compressed video. It has only little detail and it's blurry, but it looks 1000 times better than any game graphics. It proofs we can remove a lot of things.

Removing things is fine for videos where everything is batched, nothing is going to change suddenly.

So in a movie you can spend a few minutes compressing data, because you will only do it once. In a game textures have to be compressed at real time and uncompressed when used, so a heavy compression like the ones used in videos would reduce performance not gain it.

Agree for the moment, but i assume things will (and have to) change a lot to improve performance of both artists and hardware.

A good analogy is the use of scanned real world data: This data requires automatic preprocessing and detail reduction anyways which already does forms of compression similar to video. It works, creates geometry and texture data totally different than a human, and still the result is usually more realistic than human made art. (Yes i know there is still human editing involved in any use of real world data)

I know the opposite is true as well and processing human made art is much more difficult and will introduce some unwanted effects, but we'll deal with that. We already do when baking hires to game asset and no one complains, but if we automate this entire process (which is very hard to achieve), there are lots of advantages for the Artist:

You don't care about generating low poly geometry / LODs.

You don't care about creating UVs (you only use them to paint your hires model, but layout and seams don't matter)

You don't care about any kind of seams at all never again.

Could i - under the assumption that minor details get lost but overall graphics quality actually improves -  convince you as an artist with these promises? Serious question - unfortunately i have to create such a tool, so i'd really like to hear more thoughts and doupts to prepare for resistance  

Advantages for the hardware: Much less triangles, less materials, less overall fragmentation of things. Better utilization of memory (e.g. texture space). Opens up the possibility to use other data structures to represent geometry (In my case a tree of samples covering any surface for realtime GI. Basically i need only a seamless lightmap texture atlas for this, but i think about adding more features to the tool).

Warning the spoilers are boring , they are just there for people who want to know more about my opinion.

On 9/29/2017 at 7:50 AM, JoeJ said:

but if we automate this entire process (which is very hard to achieve), there are lots of advantages for the Artist:

It's true the way of the future will have artist just capturing what they need to produce there art and some kind of mind scanner will produce what ever they can imagine. I absolutely with no doubt believe this will happen, it was not what I argued against.

The two points I want to make is this:

Removing details from a mesh is not a very efficient way of optimizing models in the long term. The reason is simple it's not cost effective.

The artist should care about how the tools making the art works.

On 9/29/2017 at 7:50 AM, JoeJ said:

this entire process (which is very hard to achieve), there are lots of advantages for the Artist:

You don't care about generating low poly geometry / LODs.

There are 2 LODs a artist makes by hand this is The LOD1 and the last LOD. No matter how aggressive the LOD tool is this is how it's done. The reason is simple, it allows control over the generated LODs because you set the max and the min this way.

Having tools that are better reducing polycount won't change this, the artist needs a way to estimate the end results as they need to produce textures and shaders/materials for the LODs long before they actually have the LODs.

On 9/29/2017 at 7:50 AM, JoeJ said:

You don't care about creating UVs (you only use them to paint your hires model, but layout and seams don't matter)

You don't care about any kind of seams at all never again.

A texture isn't relative to a geometry and even if you did do this you would exceed the other limits of rendering too fast.

So to prevent the total destruction of mankind you should find a way to abandon the concept of textures, or storing data because that is what textures are, and calculate every thing at real time by creating a computer as powerful as the world; except now there is no need to reduce the geometry because it's all particles now.

So all jokes aside there is a reason this hasn't be done before, because no one can think of a way to do it.

On 9/29/2017 at 7:50 AM, JoeJ said:

Could i - under the assumption that minor details get lost but overall graphics quality actually improves -  convince you as an artist with these promises?

There is a way to do this, it's why the toon style was invented. Animators needed to reduce details and keep things as interesting as it was before. So they simply removed the unimportant details and exaggerated the important ones.

If you really do plan to make software that will reduce polycount then start here and learn 3D modeling. If you have questions about 3D modeling you can ask in the art forum. Most 3D modelers are happy to bore other people with what they know.

... much more resistance than expected, and for very different reasons.

I'll think about it, thanks!