How To Combine Normal Maps Correctly (Adding Texture Detail, Blender Tutorial) 

Yea, the comparison at the start of the video is a bit short. blog.selfshadow.com/publications/blending-in-detail/ Here is a great article comparing different methods. Their own method is exactly what I used. (They are just using quaternion rotation to make it computationally far more efficient)

@@georg240p Heidewitzka, you're fast! Thank you for the link :D

I'm guessing you're talking about Reoriented Normal Mapping method talked in that blog, this video does the exact same? I use a Photoshop version of that and it looks imo better than what even Substance has.@@georg240p Was wondering if you could make a video with the more computationally efficient way of doing it to save system resources.

Because Blender Devs. Sadly any proper fix requires some smart and persistent user to go and implement it on their own and apply for getting it added to Blender core, if the blender devs accept it or not is the second gamble issue with blender and contributions.

Although you need to do it only once and then reuse it, it's a shame blender doesn't have something similar to substance painter's normals add blend mode

@@redi4ka951 Yeah it's a crime the material menu doesn't have a build in function for combining as many normal maps as you want with just a single node group or really just that with any sort of texture.

@@RachelDetermann It will work, but you have to make sure that the actual combination is done with both vectors being in tangent space and in range [-1....1]. Not sure how blender handles this in different versions, but if you get normal data from within blender and don't export it as an image, you have to skip the range conversions since this only applies to images (typical image file formats can only store data in range [0...1] ).

@@applethefruit You can scale, rotate, translate, skew, warp the input textures by manipulating their UV coordinate systems. In my video, the Texture Coordinate Node provides the UV coordinates and this is then plugged into the Image Texture Node which samples the color of the texture at the specified coordinates. So to modify the coordinates you have to place nodes before the Image Texture Node. A simple "Mapping" Node will do the job i think.

@@georg240p yeah I did the mapping mode it worked a charm. Thanks mate.

I'm not sure if the end result is the exact same, but there's quite a difference if you compare it to for example what Substance Painter does. Normalizer washes out quite a lot of detail in comparison.

can you elaborate ?

I thought about making a video about that, here are some ideas: In case you performed a simple euler rotation: Let's say you rotated the landscape 30deg around X axis, 50 around Y and then 80 around Z. To get the tangent space normals you have to apply the inverse rotation to the normal vectors: First rotate them by -80 around Z, -50 around Y and -30 around X. In case you tilted the landscape towards a specific vector, you could actually use the node setup from this video. Just replace the base normal map by a constant vector (the vector by which you tilted the landscape) But this would just apply the same tilt again so you have to do the inverse tilt (just flip the sign of the X and Y coordinates of the tilt vector) In case you have no idea what rotation you performed, you can always convert between two coordinate systems if you know where the X,Y and Z axis of the coordinate system ended up after the rotation. To put the world space normals into this rotated coord system (tangent space), just project the normals onto these 3 axis. The tangent space X coordinate is just dot(worldNormal, XaxisVector), Y = dot(worldNormal, YaxisVector), Z = dot(worldNormal, ZaxisVector) This is basically just applying an inverse matrix by hand. Feel free to message me on discord if you want to send screenshots or something: umsoea#8675

It should work for combining any two (or more) tanget space normal maps. I added a download link to the my final file in the video description.

The link to the last video is in the video description. You might have to reload the page. vecToSpherical converts a 3d vector to spherical coordinates. I showed how to create it in the last video.

6:15 By rendering an image and saving it (if you followed my scene setup) A link to my final .blend file is in the description.

At 5:40 I showed how to change the strength of the detail normal vector (by manipulating the angle theta). You can do the exact same thing to the base normal vector (bottom one). This way you can control the strength of both normal maps independently. And you can also do the same thing at the end (after the combination) to change the strength of the final result. You can combine as many normal maps as you want, and change the strength at any point in between. Not sure what you mean by "the nodes branch away from the multiply node". (The multiply node is only used to manipulate the angle theta). In case you want to send screenshots of your node setup, feel free to message me on discord: umsoea#8675

Works perfectly fine for me. In the video you can see that the base normal map also has a flat background. In which way does it break for you? The arctan2 function might cause some trouble. Check if it returns something close to zero for phi. If not, set phi to zero IF the z component is close to 1. Here is my file in case you want to double check the node setup: www.mediafire.com/file/y60i9en2qltabyo/tut12_combine_normal_maps.blend/file (i used blender 3.3) If you need a faster and numercially more stable version of this approach (using quaternions), Stephen Hill has a great blog post: blog.selfshadow.com/publications/blending-in-detail/

You're right, It's a bit confusing especially because I said that we are rotating by theta and phi. ("2 rotations") Here is what we actually want to do: We want to perform a single rotation (by theta) BUT around a rotated axis. And this rotated axis is the X axis that has been rotated by phi around Z. The problem with that: With simple Euler rotations, we can only rotate around the main axes: (X,Y,Z) But we can use a trick: If we first align the rotation axis with one of the main axes (X,Y or Z), we can just rotate around this main axis (by theta). Because they are identical. We just have to make sure to reverse this alignment at the end. So here are the 3 steps: 1. Align rotation axis with one of the main axis: As mentioned above, our rotation axis is just the X axis rotated by phi around Z. If we reverse this, the rotation axis is identical to the X axis. So we rotate by negative phi around Z. 2. Perform the actual rotation. Since the X axis and our rotation axis are now identical we can just perfom a rotation around X by theta. 3. Reverse the alignment. Rotate by positive phi around Z. Now the rotation axis is back to where it was at the beginning. This is what I showed in the video with the example of rotating the cube. Hope this helps.

Shouldn't you just be able to do a single quaternion rotation? The quaternion for the rotation in question is just sqrt(final / initial).

Stephen Hill from Lucasfilm has a great blog post comparing different techniques: blog.selfshadow.com/publications/blending-in-detail/ You seem to describe the first technique they mention: "Linear Interpolation" (Vector addition, which is the same as averaging the two vectors). The technique I used in my tutorial gives the exact same result as their proposed method. (They use quaternion shortest arc rotation to make it computationally extremely efficient, but hard to understand)

just use the resulting vector as the new base normal vector for the 2nd combination

why not just use quaternions? numerically stable and the most efficient.

@@georg240p Vectors and matrices are the simplest and cheapest (computationally)

@@xbzq The quaternion version I tend to use requires 1 divison, 6 multiplications, and 6 additions. and can be used to rotate points aswell. Stephen Hill has a great blog post about it: blog.selfshadow.com/publications/blending-in-detail/ Would love to know if there is anything faster. Keep in mind we are doing a shortest arc rotation.

​@@xbzq Quaternions are cheaper to compose and store than matrices, but more expensive to apply. As for which is simpler, quaternions are only "complicated" because they're obfuscated and communicated terribly. They're really just a blending of a 180° rotation (represented as a blending of 180° rotations around each axis) and a 0° rotation.

In newer Blender versions (eg 3.6) there is the Mix node (set to float). Set the mask as the Factor input. And then you can control A and B separately. Use its output as the normal strength. In older Blender versions, just use the MixRGB node. Does the same thing but it looks odd because its using color values.

@@georg240p in case if we use black and white texture because it has to be exact location for label/box etc? Will it work or I’m just connecting it wrongly? I also face a problem where it’s black and white but when I change slots the white turn grey instead of completely black while before the black part was completely black. Like it inverted but there’s no completely black in the inverted but grey even tho I change another slot color to completely black. Do you have any suggestions?

@@natsunwtk Yes, to get a clear separation, the mask should be only black and white (0 or 1). Im not sure what you mean by inverted or gray. You mean the mask looks wrong when you set it as the output? Try to set the Color space to "Non-Color data" in the image texture node that loads the mask.

@@georg240p yes, something like that. It’s already non color when it’s before inverted it showing completely black and completely white but when I change the slot A to slot B in a mix color node, it showing white->gray and black-> white. Sorry my english is not good enough to saying it properly. (I connected a black and white image to a mixed color node for bumping. It’s mask just fine because I also used it for roughness. I just connected them to checking if it’s masking properly but turn out it appeared gray instead of black for bump node)

​@@natsunwtk The result of the mix node are the strength values that will be used to control the normal map strength. It is not a mask! If its inverted, thats because you defined the colors this way in the mix node. And maybe thats exactly what you want. Connect the output of the mix node to the value that you previously used to control the strength of the entire normal map (eg in my video I used a multiply node. So in this case connect the mix node output to the bottom slot of the multiply node)

if you like it, why not?

​@@georg240p well, that's an option too

Should work fine to just use the resulting normal vector from the first combination as the new base normal vector. just make to keep the z component positive +renormalizing.

@@georg240p got it, thank you very much!

No, normal maps store 3d vector data and treating them as color will 1. produce unnormalized vectors (a normal vector should always be normalized). and 2. it doesnt follow any geometric concept so results can be quite unpredictable. For more details, here is an article explaining this in more detail by comparing color blending techniques with the geometric approach: blog.selfshadow.com/publications/blending-in-detail/

@@georg240p Thanks for info. I'm new to all of this.

I tried it. It works but result is kind of dull.

bruh

@@georg240p how u been 😂

pssst

then just use ovelay. who cares. This technique is for those not satisfied with regular image blending techniques. I mentioned that in the first few seconds of the video.

It works, but it's not the "mathematically" correct way of doing it. It washes out a lot of detail, but if one doesn't care about it, it doesn't matter. But I personally would prefer to have both of my textures looking as best as they can and how they are intended.