Home 3D Art Showcase & Critiques

Tutorial - Gradient Mapping & UDK

This week on Idol Gives Back ;)

Part 1: Generating color textures from heightmaps in PS
Part 2: Doing the same thing in a realtime shader in UDK
Part 3: More crazy ideas.

GradientMapping_01.jpg
GradientMapping_02.jpg
GradientMapping_03.jpg

So, a bit rushed but I wanted to get this quick tut out there, I hope it atleast gets some of you guys' creative juices going. A quick runthrough of some of the stuff I've been fiddling with and thinking about lately - be it screenspace, texture space, by lightvecor, etc etc, mapping things in interesting ways is just one of those awesome things we can do these days. I'll try to amend it to show a hipoly source as well as other things, just didn't have more time this morning. By all means keep your ideas coming and feel free to clutter up the thread with crazy attempts :)

Thanks guys!

Replies

  • glynnsmith
    Offline / Send Message
    glynnsmith polycounter lvl 17
    This is neat :D I'm sure something similar to your method was used by Valve, in Source, to get more variants out of their greyscale zombie textures.

    You can also add a gradient to alter the colour/saturation/etc as your scene transitions from lit to shaded.

    Cheers for sharing.
  • Xoliul
    Offline / Send Message
    Xoliul polycounter lvl 14
    This sounds very interesting, but I'm not sure I get everything.
    Part 1 is about how you use this as a sort of procedural basis for your diffuse textures, having them generated based upon a heightmap ? Purely a photoshop thing ? Mind showing how you do this exactly, it isn't really mentioned and I haven't used this before.
    I can see how it looks pretty good for the tree bark, but I can't really imagine how you come up with that specific gradient to attain such a look. You have a library of these things or something ? Mind showing some more examples ?
  • glynnsmith
    Offline / Send Message
    glynnsmith polycounter lvl 17
    White mask value = left of the gradient and black mask value = right side of the gradient (or the other way around, I can't remember). I think.

    So, it pulls values from the gradient where the mask transitions between black and white.
  • rasmus
    Glynn: Thanks man, and yes, that sounds about right

    Hey Xoliul, yeah, that's the part where it's a bit rushed, but yeah, basically part one is just Photoshop stuff - you do a gradient map on top of other grayscale layers, found under Adjustment Layers, then just go nuts with the color sliders and add whatever gradients you like to your library. It's really just a matter of playing around to get what you like and iterate on it to get a base texture that is in line with the heightmap/bumpmap/normalmap without having to paint colors by hand. I'm big on that *P I'll try to come up with more examples when I have the time

    I'm also big on trying to find ways of getting at "sub-pixel" stuff in textures - ie, stuff that happens as an interpolation result between two or more other factors to get away from a set texture resolution, or purely as a style experiment. So that's part 2 :P
  • glynnsmith
    Offline / Send Message
    glynnsmith polycounter lvl 17
    Not sure I really helped explain anything that Xol wouldn't already know, with my previous reply :P

    But I think the question is: How do you visualise which value of your mask will pull which specific value from your gradient?

    I'm unfamiliar of any method in PS that'll let you "colourise" a greyscale mask with a gradient, like that.
  • rasmus
    Glynn: It's all realtime as you edit your gradient map in PS, try it and you'll see
  • glynnsmith
    Offline / Send Message
    glynnsmith polycounter lvl 17
    You beauty!

    Re-read your reply, and realised Gradient was different to Gradient Map, which I'd never looked at in all my years of using PS.

    What a handy thing to know! Thanks man :)
  • rasmus
    I know, I spent way too many years in ignorance of it myself :) There is one other thing you can do though which I did before switching to gradient maps, which is to manually Select > Color Range and pick a range out of the grayscale base, and use multiple filled selections like that in layers to build up your texture - all in all though Select > Color Range is very useful for specific selections but not as nice as the gradient mapping for this thing.
  • Xoliul
    Offline / Send Message
    Xoliul polycounter lvl 14
    Makes more sense now. And Glynn, yes I got that part, I've seen that before with TF2's ramped shading, but this is being used for diffuse colors instead of shading colors.
    So rasmus, do you finish the greyscale "depthmap" completely before you start creating the gradient? Or do you go back and forth a bit?
  • Vailias
    Offline / Send Message
    Vailias polycounter lvl 18
    And a Q3 era texturing technique comes back to life. :)

    Sounds like there needs to be a dedicated photoshop tips n tools section of the wiki or some other central repository.

    Also my guess as to the need for adjusting the RGB curve to match PS is PS's working gamma/colorspace being SRGB or AdobeRGB while UDK's is linear. SRGB and ARGB have a gamma value of 2.2, so to reverse that shift (ie correct it if viewing in linear space) you'd need 1/2.2 or, like you found, about 0.5 (technically 0.45... but hey.)
    That's actually really good to know.. thanks. :)
  • rasmus
    Xoliul: The nice thing is it stays on top and works with whatever happens underneath, so I find myself adjusting both the gradient and the heightmap layers in tandem. Like just now I duplicated the Gradient Map a few times, and every time it works with the results of the previous gradient, causing automatic variations that you can switch off and on. Stuff like that. But whatever works, really! I think the more techniques you can mix toghether the more interesting it is.

    Like I said, it would be fun if people posted any other crazy shader/texture things they might come up with, gradients or not. No such thing as mixing too many crazy ideas :)
  • rasmus
    Vailias: Yeah I figured the RGB curve was something to do with that, thanks
  • Envart
    Offline / Send Message
    Envart polycounter lvl 6
    I really like the postprocess shader. It looks kinda painterly. Would it be possible to have warmer tones closer the camera and the cooler tones in the distance? Nice work.
  • rasmus
  • [HP]
    Offline / Send Message
    [HP] polycounter lvl 13
    This needs to be in the news! Uber interesting.
  • d1ver
    Offline / Send Message
    d1ver polycounter lvl 14
    This is effing fatnastic, Raz you sexy beast! Just tried it out and it works like a charm.

    This could be really useful on the tileable/organic stuff. Too bad it doesn't eliminate a diffuse texture lookup, even though it's small. You could make a gradient procedurally jumping through some hoops in UDK, but if epic made a gradient node this would be priceless. Btw is it possible with a custom HLSL node?

    Once again, amazing technique. Can't wait to try it out in production.
  • rasmus
    HP: Thanks dude, knew it would tickle a certain slice of the population here :)

    d1ver: Thanks - you're right, and as I said, it lends itself best to tiling organic stuff for the PS texture creation, shader-wise the possibilities are endless. I don't have a problem with the texture lookup so far, it basically eliminates any visual traces of texture compression etc, but yeah, I have yet to try out the UDK stuff on a larger scale and for sure it is more of a style experiment than a replacement for normal texturing. The thing with offsetting UVs per-pixel is that sooner or later everything starts looking like melting icecream :P I'm sure you can do this with any shader language, offsetting UVs per-pixel is a common feature of SM3 as far as I can tell, it doesn't really matter what you do it with. Definitely one of the more exciting aspects of this generation graphics tech though :)
  • Vailias
    Offline / Send Message
    Vailias polycounter lvl 18
    The only issue with procedurally gradient mapping in udk atm is the number of lerps you'd wind up putting in for more than a single color change.
    if you want a 2 color "gradient node" just make 2 color constants and lerp between them with the alpha being whatever your were gradienting over.

    You could do it in a custom node programatically, but its would not be easily changeable at runtime, or you'd need n number of float4 inputs to the custom node.
    Use the RGB channels for color, and the A for position along the overall gradation, then you can iterate over the gradient, and selectively lerp between each color pair based on value of the input pixel.

    edit:
    if you want to replicate the texture to texture example that's less code to deal with than my above example.. and I didn't feel like writing it all out. :)
    here's the HLSL for the equivalent node setup in udk
    float3 output;
    float2 UV = float2(ControlTexChannel,0);
    
    output = tex2D(Gradient,UV);
    
    return output;
    
    
    HLSL_GradMap.jpg
  • rasmus
    Vailias: That's interesting, but what would you say the benefit to this procedural method is?

    EDIT: As a start to a proper gradient node it's cool - you'd need an interface though! :)
  • Vailias
    Offline / Send Message
    Vailias polycounter lvl 18
    You would just question me on just enough to get me to write it wouldn't you?
    :P Thanks for the push. It took far longer than I'd thought all because it took me a while to remember/figure out the general form for making these sub gradations, but the mental math felt good. I finally had to break out the paper and work out what I needed in steps to find the formula. Its simple to the point that I'm going to blame lack of sleep for it taking so long. ;) anyway on with the content..

    Yes you would need an interface for it to be artist friendly. And a proper gradient node with unlimited color stops would need to be coded in C++ with access to the full unreal source.. or well at least the source that pertained to shader nodes.

    One benefit to having it be procedural is you get much finer color resolution and control than by using a texture map. This is basically moot if you use a texture map for a control source, but if you apply it to a floating point value like the depth buffer, you get infinitely smooth transitions. Bilinear filtering for textures only does so much.
    eg
    HLSL_GradMapProc1.jpg

    The other, and to me more tangible, benefit of doing gradient mapping procedurally, is the colors could be set to be parameters, accessible by script, so you could change any given color and even its placement on the gradient on the fly in game.
    Think of things like changing ranges of IR vision, or even faking entire different spectrum effects. Yes AVP did this, I don't recall how, though it very well may have been a selective screenbuffer effect much like this programatic approach could be applied to.

    Another thing to note with either of these approaches, texture or code based, is the interpolation can be from ANY floating point value, be that depth buffer, single channel of a texture, or something like time, or a trigonometric output. You may notice some unused nodes in the below screenshot. One of those is hookedup to a modifiable sine node attached to a time input. So what it does is smoothly ramp through each of the colors in the gradient and back down, forever. Basically creating a multi step blink/fade effect.

    Another benefit is not having to use an additional texture sample in your shader at the expense of some code. If you already have expensive use of texture memory, this may work better. If you have free memory but you are fill rate bound, and or have heavy weight shaders already going on, then another texture might be simpler.

    I also have to state that epic mentions using the custom node may make less optimal shaders than their node based system, as the custom node, and I believe the entire shader it is in, cannot be evaluated by their shader optimizations.

    Here's a general HLSL setup for a procedural gradient map.
    float3 output; //output color
    
    const uint NUMCOLORS = 4; //number of colors for programatic gradient
    
    float4 gradient[NUMCOLORS]; //here's the actual gradient keypoints
    gradient[0] = color1;
    gradient[1] = color2;
    gradient[2] = color3;
    gradient[3] = color4;
    
    /* This method has a flaw, in that the colors listed here are essentially required to be listed in ascending order of alpha value, else the blending will not be smooth between adjacent colors.*/
    
    float4 blend1 = gradient[0];
    float4 blend2 = gradient[0];
    float3 weight;
    float singleWeight = 0;
    
    for (int i = 0 ; i<NUMCOLORS-1; i++ )
    { 
     if (gradient[i].a <= ControlTexChannel) //if the position of our gradient key is equal to the control value
      {
    
        blend1 = gradient[i];
    
        if (i == NUMCOLORS-1)
       {
        blend2 = gradient[i];
        }
        else
        {
        blend2 = gradient[i+1];
        }
       }
    }
    
    singleWeight = (ControlTexChannel- blend1.a)/(blend2.a - blend1.a);
    weight = float3(singleWeight,singleWeight,singleWeight);
    
    output = lerp( blend1.rgb, blend2.rgb, weight);
    
    //output = color3.rgb;
    return output;
    
    HLSL_GradMapProc2.jpg
  • d1ver
    Offline / Send Message
    d1ver polycounter lvl 14
    Hey, Vailias, this is totally awesome. A lot better than messing around with all the lerp and uv nodes. And thanks for the comments - made the thing more transparent for the unfamiliar with HLSL. I learned a thing or two about it thanks to you.

    The color parametrization is also useful to just quickly create a bunch of instances.

    I'll throw it into UDK to check it out.
  • d1ver
    Offline / Send Message
    d1ver polycounter lvl 14
    Hey man, I thought I'd help you out a bit with a sorting function so you don't have to stick those color nodes in a particular order. Just the alpha values set up properly.

    gradientmappingsorting.jpg
    float3 output; //output color
    
    const uint NUMCOLORS = 4; //number of colors for programatic gradient
    
    float4 gradient[NUMCOLORS]; //here's the actual gradient keypoints
    
    gradient[0] = color1;
    gradient[1] = color2;
    gradient[2] = color3;
    gradient[3] = color4;
    
    //Sorting Function
    
    float4 TempColor;
    int StillSwapping=1;
    
    while ( StillSwapping )
    {
     
      StillSwapping = 0;
    
      for (int i = 0 ; i<NUMCOLORS-1; i++ )
      {
    
         if( gradient[i].a > gradient[i+1].a )
           {
    
           TempColor = gradient[i];
           gradient[i] = gradient[i+1];
           gradient[i+1] = TempColor;
           StillSwapping = 1;
    
            }
      }
    
    }
    
    //Sorting End
    
    float4 blend1 = gradient[0];
    float4 blend2 = gradient[0];
    float3 weight;
    float singleWeight = 0;
    
    for (int i = 0 ; i<NUMCOLORS-1; i++ )
    { 
     if (gradient[i].a <= HeightMap) //if the position of our gradient key is equal to the control value
      {
    
        blend1 = gradient[i];
    
        if (i == NUMCOLORS-1)
       {
        blend2 = gradient[i];
        }
        else
        {
        blend2 = gradient[i+1];
        }
       }
    }
    
    singleWeight = (HeightMap - blend1.a)/(blend2.a - blend1.a);
    weight = float3(singleWeight,singleWeight,singleWeight);
    
    output = lerp( blend1.rgb, blend2.rgb, weight);
    
    //output = color3.rgb;
    return output;
    
    

    My only worry is that there's no way of checking if two colors have the same alpha value. That could create a bit of misunderstanding(and subdivision by zero). I just don't know the code in HLSL that could give out a warning or smth.
  • glynnsmith
    Offline / Send Message
    glynnsmith polycounter lvl 17
    One thing I've wanted to write, just so people that haven't thought of it can get a decent idea of how far this can go (and why I was so excited about it when rasmus posted it up), is splitting up a single gradient into multiple ranges, to allow for colour usage variants - Valve did something similar with the L4D2 gradient mapping for their generic zombie assets.

    As a really basic example, you have a character that you're about to fit with a grey-scale mask and a gradient map to colourise it.

    On your texture mask, you set your skin to values between 0-25 brightness, the trousers would be between 26-50 brightness, the shirt would range from 51-75 brightness, and the shoes would go range between 76-100 brightness values.

    So you split your gradient into four quarters. On the gradient ramp, you set a range of 0-25 for the colour of the skin, 26-50 for the trousers, 51-75 for his shirt, 76-100 for his shoes.

    This allows you to set multiple colour sets to different parts of your model/mask, and should allow you to get more out of such an efficient way of texturing your assets.

    @Vailias - That's pretty sweet, man. Could imagine being able to translate gradient settings to that, with little problem. Very cool :D
  • feanix
    Offline / Send Message
    feanix polycounter lvl 7
    This is the best thread ever. You are all awesome. My puny artist brain stands in awe.
  • Vailias
    Offline / Send Message
    Vailias polycounter lvl 18
    D1ver: Thanks. I was about to write in commentary that the ordering issue could be solved with a simple bubble sort like you did. I didn't put it in though as the whole of that HLSL block gets evaluated for EVERY PIXEL of the material at EVERY draw call its visible. On a simple thing like this it doesn't add that much overhead, but if we were to use more than 4 color points, in a complex material setup thats a lot of extra comparisons per second for things that may not ever change.

    The simplest solution is to have a variable version and a static version and only use the variable version for assets that absolutely need it in the game.

    I'd also thought of and notice the divide by zero potential in my code. It would be a matter of throwing another "if" statement or ternary operator in there somewhere to check if the alpha values were identical before setting up the subtraction.

    Where I have it it winds up being .. farily safe mostly thanks to the compiler I belive. The only case in my initial shader where you can divide by zero is where you get a 0/0 situation. This still should cause some undefined behaviour, but I'm gathering that UDK's compiler has some safeguards to evaluate that to 0 rather than throw an exception.
    Another method to fixing it would be multiplying 1/((Heightmap - blend1.a)+0.00001) * (blend2.a-blend1.a). So then you wind up multiplying by zero if the weights are the same, which is fine because a zero weight will always return the starting color.
  • rasmus
    Great stuff, guys! And ofcourse you make some great points as for what a true gradient node could do - in my mind though, it would need to offer unlimited colors on the gradient to compete with Photoshop for doing certain things, but to be sure, this is great work and a big step on the way. D1ver, I don't know how you do the sorting, but it's cool :) And Glynn, you're absolutely right - those are the kind of things that really rock.

    I have to say though Vailias, I can't seem to get those ugly blocky compression artifacts to show up as a result of using a texture lookup - even with it compressed to Default as opposed to Grayscale. True it's not free from a little banding:

    TexArtifact.jpg

    I'll try to post some more examples over the weekend, hope I have the time because this thread is rockin ;)
  • rasmus
    Just threw together some more PostFX shenanigans, could be quite interesting:

    GradientMapping_04.jpg
  • Vailias
    Offline / Send Message
    Vailias polycounter lvl 18
    :) Just had to play around some more with this and check out how it could work for a post process shader. This one uses a desaturated scene sample with a bit of green boost in the desat, plus the "uber post process effect" inline after it for things like bloom, grain, and dof.

    All these come from the same code I had before, the 4 step color thing, with the colors changed into parameters so I could easily tweek them in a Material Instance while the scene was running.
    I of course went directly to the nightvision effect, then played with some non realistic techniques to see how they'd look on a level not built for them.
    Good times.
    MapPostShader_1.jpg
    MapPostShader_2.jpg
    MapPostShader_5.jpg
    MapPostShader_3.jpg
    MapPostShader_4.jpg
  • Envart
    Offline / Send Message
    Envart polycounter lvl 6
  • rasmus
    Ah, see, there we go :)
  • ozzmeister00
    Vailias got me thinking. The other day I put this together in UDK: goo.gl/jNsua

    While monkeying around with it I lamented that iterating on the rainbow texture ate a fair bit of my time. Meanwhile a friend had showed me this thread and I put two and two together and went from this:

    01_texture.jpg

    This this (I append the RGB output of the VectorParams with its Alpha output to get the right float4 input the HLSL is looking for. If anyone has a better way, I'd love to hear it!)

    03_append.jpg

    So I could do this:

    06_instance.jpg


    Documentation Here

    All I really needed to do was add more entries to the Gradients array.
    //code authored by d1ver and Vailias on polycount
    float3 output; //output color
    
    const uint NUMCOLORS = 8; //number of colors for programatic gradient
    
    float4 gradient[NUMCOLORS]; //here's the actual gradient keypoints
    gradient[0] = color1;
    gradient[1] = color2;
    gradient[2] = color3;
    gradient[3] = color4;
    gradient[4] = color5;
    gradient[5] = color6;
    gradient[6] = color7;
    gradient[7] = color8;
    
    /* This method has a flaw, in that the colors listed here are essentially required to be listed in ascending order of alpha value, else the blending will not be smooth between adjacent colors.*/
    
    float4 blend1 = gradient[0];
    float4 blend2 = gradient[0];
    float3 weight;
    float singleWeight = 0;
    
    for (int i = 0 ; i<NUMCOLORS-1; i++ )
    { 
     if (gradient[i].a <= ControlTexChannel) //if the position of our gradient key is equal to the control value
      {
    
        blend1 = gradient[i];
    
        if (i == NUMCOLORS-1)
       {
        blend2 = gradient[i];
        }
        else
        {
        blend2 = gradient[i+1];
        }
       }
    }
    
    singleWeight = (ControlTexChannel- blend1.a)/(blend2.a - blend1.a);
    weight = float3(singleWeight,singleWeight,singleWeight);
    
    output = lerp( blend1.rgb, blend2.rgb, weight);
    
    //output = color3.rgb;
    return output;
    
  • daufoi
    is anyone else having trouble viewing the 1st image of the tutorial?
  • rasmus
  • Fingus
    Offline / Send Message
    Fingus polycounter lvl 11
    Hello everyone, I've been experimenting with gradient mapping in game engines and I found this thread.

    This is an extremely simple solution I came up with to the issue people were talking about with having several gradient maps on a single asset without having to rely on complex lerping. The shader setup shown here is literally just the grayscale into the gradient map UV's.

    4way_gradientmap.png

    I simply used Photoshop's Gradient Map tool and assigned different grayscale ranges across the texture, and then stacked four different gradients in a row. Mipmapping will have to be disabled on the gradient to avoid color bleeding, and the smaller grayscale range per gradient will likely impact quality in some way, but for simple assets it should be passable.
  • rasmus
    Thanks for sharing, Fingus - long live the gradient thread :)
  • LoTekK
    Offline / Send Message
    LoTekK polycounter lvl 17
    You'll want to have a small buffer range between each gradient, though, to account for compression artifacts.

    My only issue with that is the relative lack of flexibility in the gradient maps themselves (you'll need a different set of gradient textures per color set, unless you start playing with stacked sets of gradient maps).
    I wrote this post some time ago, and it's not UDK-specific (though the shader code is more than doable with material nodes), but it covers some general ideas:

    http://teckartist.com/?p=26
  • Fingus
    Offline / Send Message
    Fingus polycounter lvl 11
    I don't have UDK here right now so I can't test it, but perhaps something like this?

    LhMGs.png

    Not 100% on the if statements, because I suspect the bottom grayscale values will cause problems. I suppose it could be fixed with another if statement to cull out the bottom range. In either case this shader network suddenly got rather lumpy again. How expensive is a Lerp anyway? I have the impression they are somewhat costly, but I'm not sure if they're bad enough to warrant a workaround like this.
  • Nielsen
    Offline / Send Message
    Nielsen polycounter lvl 8
    Vailias wrote: »
    The only issue with procedurally gradient mapping in udk atm is the number of lerps you'd wind up putting in for more than a single color change.
    if you want a 2 color "gradient node" just make 2 color constants and lerp between them with the alpha being whatever your were gradienting over.

    You could do it in a custom node programatically, but its would not be easily changeable at runtime, or you'd need n number of float4 inputs to the custom node.
    Use the RGB channels for color, and the A for position along the overall gradation, then you can iterate over the gradient, and selectively lerp between each color pair based on value of the input pixel.

    edit:
    if you want to replicate the texture to texture example that's less code to deal with than my above example.. and I didn't feel like writing it all out. :)
    here's the HLSL for the equivalent node setup in udk
    float3 output;
    float2 UV = float2(ControlTexChannel,0);
    
    output = tex2D(Gradient,UV);
    
    return output;
    
    
    HLSL_GradMap.jpg

    I tried to give this a whirl in UDK because it's really cool but I stumble upon this if I use the exact setup: "error X3014: incorrect number of arguments to numeric-type constructor"

    I saw this:
    2kaEpEv.jpg

    and this opens up some opportunities with regards to fast workflow / optimisation in UDK.
    Image is by bungie, you can get the full talk on Bungie VFX for Halo3 at
    http://halo.bungie.net/Inside/publications.aspx

    Any idea what I did wrong? Am I missing something ?


    -- FIXED! I used regular texture instead of CustomTexture...
  • Omnicypher
    Offline / Send Message
    Omnicypher polycounter lvl 9
    Gradient maps can be used for more than just diffuse and emissive.
    Im Making a shader that can blend between rubber, plastic, metal, crystal, and emissive material properties, based on a gradient map. i just stack gradients that represent masks for each material property. i plan on using this to define spec, gloss, emissive, fresnel glow, cube map reflections, opacity, transparency, and metalicity. Using this method, i can have alot of interesting material variety without wasting entire maps on rare things like glow or transparency.

    [ Images Deleted due to photobuckets greed. ]
Sign In or Register to comment.