Request - Normal Map Comparison (Baked Vs Generated)

Del

~ Okay, I'm bored of looking through the wiki, and the sticky'd threads now. I can't find what I'm looking for I'm hoping I could get some help?

Are there any images of normal maps that are generated from flat images (e.g. crazybump normal from a photo of bricks) compared to normal maps that are baked from a high poly. I know that recently studios are pushing towards baking everything because the results are much crisper/richer/accurate, but I just want to see a comparison for myself.

Thanks V much (Hope this is the right section for this request)

Joao Sapiro

its simple , just make an HP piece that is slightly complex , do its lowpoly generate normal etc, then use that LP and try to have the same shading with crazybump generating normal

Del

~ I just wanted to see if someone had done it already... Someone who's good at all this crazy 3D stuff

SHEPEIRO

TBH for me its not an issue of quality (3d generated will be better) more a question of style... photosource gives a (admittedly slighly unccanny) realism, whilst hand sculpted will always have some hand crafted look to it... i dont think that will ever change

Will Faucher

It depends what kind of normal map you want too, though. Making, say, a pillar with lots of inctricate details, wear, and, say a dragon/snake wrapped around it, a generated normal map would look terrible. I find generated normal maps have a hard time defining shape and volume. Crazybump's "enhance normal map volume"/shape recognition helps this, but for hard-surface stuff, it's pretty bad.

But for rocks, and sometimes, bricks, (most organic stuff) generating the normal map from crazybump yields very nice results.

Ben Apuna

I think it's really the difference between per-pixel VS. per-vertex lighting.

The normal map is there to compensate for the lack of vertex resolution in a typical low poly model. This is really visible in the lack of gradients in a typical 2D generated map. The gradients help compensate for the lack of verts that would be needed display the model's desired form and shape.

If you use normal maps generated from a high poly model then you get all the good lighting information from it on how the surface interacts with light for every pixel you have on the texture.

On the other hand with a 2D generated normal map your low poly model is stuck with using vertex normals (smoothing groups/hard-soft edges) to calculate the way light moves across the surface. Admittedly, you can do a LOT with vertex normals, but you start with a much lower resolution of lighting information to work with. Because you then have to paint or photoshop in all the missing surface lighting information. Which IMHO is much harder to do convincingly than to just model the high poly in the first place.

Oniram

from the uncharted 2 art direction pdf.

Will Faucher

I was just about to post that. But I find they exaggerated a little. I took that diffuse map (just for a test), brought it into Crazybump, and got a result that is very similar to the Zbrush normal map. The difference is, I couldn't get the same crisp lines between each rocks, which is pretty normal, but the crazybump result is just overall really bad in that example. :P

Adam L. Gray

Hmm, I'd say it usually boils down to:

HP Model = Quality
Generated = Saves time

Sometimes a combination of the two is a great way to go. But it really depends on the asset, too. And art-style as SHEPEIRO mentioned. I can't confirm it, but from what I've heard, at DICE they tend not to work with hp models at all unless necessary. Then again, they use quite a specific art-style, a bit like what is commonly seen with the source engine

Cheers

Oniram

i also think now with the nDo action for photoshop, you can get even better results than crazybump. but i do agree that zbrush models are more time consuming, but i think the look sometimes is a bit more satisfying. and i guess also if the mesh is intended not to be something flat and tileable (like the rocks example), then it still may be necessary to bring it into zbrush.

Will Faucher

Ehhh I think nDo is alright for some things, but... I've found it somewhat annoying to work with. Maybe it's just me.

Adam L. Gray

Hmm, I gave it a go myself, and I'm by no way saying this is a correct representation. I haven't used Crazybump nearly as much as I've used the Nvidia filter all this time before I got CB, so need to brush up on that a bit.

But I still found the results quite interesting, and tbh, the Nvidia filter can work quite well if you combine multiple layers and overlay them on top of each other like I did here.



oops, forgot the normal-maps:



Cheers

Ben Apuna

Off topic:

The main difference between PS and Crazy Bump is not about building up volume from 2D source because both can do this well, but instead in the way they handle combining normal maps.

Crazy Bump can take two normal maps that would point light in different directions and combine them correctly using vector math. Photoshop just can't do it quite right.

Will Faucher

Tiros wrote: »

Hmm, I gave it a go myself, and I'm by no way saying this is a correct representation. I haven't used Crazybump nearly as much as I've used the Nvidia filter all this time before I got CB, so need to brush up on that a bit.

But I still found the results quite interesting, and tbh, the Nvidia filter can work quite well if you combine multiple layers and overlay them on top of each other like I did here.



oops, forgot the normal-maps:



Cheers

You're not using crazybump right. :P Boost the intensity, enhance normal map volume. You'll get much better results. Also, your shape recognition seems a little... off. And why are your normal maps green channels flipped?

Dylan Brady

I dunno I think if were using photo sourced textures we have to consider light angle, and the fact that color effects the depth when transferred over.
So the results could potentially be really off (most artists know how that looks, and avoid it like the plague)
But with baking you can be sure that the shape your baking will light exactly the same on a flat plane.

rasmus

That Uncharteed 2 pic should say it all I'm not sure I'd say you can't get photorealistic results with modeled textures either... That is up to the artist and ofcourse the style of the project. But basically the control you have over creating the model yourself versus clone-tooling photos makes it the way to go if you have the time imo, especially considering all the height, AO and cavity data etc you get for free - which is a huge plus when creating the diffuse/spec/masks etc.

One thing I find really annoying lately is seeing the CrazyBump tell-tale diagonal "slant" to some normalmaps in games. Then again I'm sure it has its uses, and Photoshop as well, for additions to modeled stuff or on its own for simple stuff.

eld

First and foremost it's important to understand exactly how normals work before deciding which one method is the best way to generate them.
Know how to use all your given tools, and you'll get great results.

I've seen fantastic artists work magic using just very little baking.

http://www.philipk.net/tutorials.html

Adam L. Gray

Prophecies wrote: »

You're not using crazybump right. :P Boost the intensity, enhance normal map volume. You'll get much better results. Also, your shape recognition seems a little... off. And why are your normal maps green channels flipped?

I know, as said, I still need to brush up on that a bit :poly124:

The intensity was more or less up at the top though, and the normalmap volume couldn't be enhanced much more in there, without it just ending up in a big pile of bump.

About the normalmap's green channel being flipped, it's not, not for max at least. Would prolly come out wrong if you were to use Maya tho

iconoplast

Just so, eld. There are some other ways to improve the results from nvidia's filter too. Doesn't work on everything, but what does?

Oniram

yeah ive had that method of overlaying layer after layer from the nvidia filter, but sometimes it starts to get that cloudy shape recognition type of look that crazybump gives.

Edit: just did a quick test using a texture i did a while ago. I used nDo, xnormal, and nvidia. i dont have crazybump anymore so couldnt use it for an example.

generally, i would say that the zbrush bake came out the best, BUT, if i were to use a combination of the nDo and xnormal filter, i think that could give better results.