Pseudowatercolour V1

I’ve nearly finished a Ryou model recently. Although, having done that, I had the problem that I didn’t have a good shader to apply to it.

I’ve spent quite a lot of time in research and experimentation for various aspects of my shaders, for fun and to use in my art, but I’ve never really got something I would say is complete and reliable. Perhaps because I was too perfectionist about it, and nitpicked a lot over 3d elements. Making 3d that’s a convincing facsimile of 2d has been a goal of mine for a long time. I always wanted to be able to draw well, but the amount of work it needed was daunting, and I got put off. It’s ironic that with the amount of effort I’ve put into NPR to imitate it, I could’ve just learned the real stuff by now.

Still, I’ve been trying to progress. I want to achieve what I set out to do this year, making more art and models, so I made a Ryou base mesh. From that, I want to derive several, like Yami Bakura, fem Ryou, au versions. With that, I needed a shader, so I went back to my previous efforts.

My previous work in steps; it may look similar, but a lot of them derive their appearance differently.

I was dissatisfied with what I had previously, though, so I considered my approach. The biggest problem was the silhouette. It was too 3D, with fresnel, the easiest real time method I knew of to get something resembling an outline, capturing too much detail. I considered editing the normals, but then, I also need a correct set of original normals for the shading. Having multiple sets baked seemed like too much of a pain. I also experimented with using nodes to blur a texture that defines the normals to try and get a blurry, and thus less detail-accurate, fresnel, but it didn’t work when viewing it from different angles. Then I came across this, a discussion on how to, essentially, replicate a depth pass with a shader.

I replicated the effect, so now I have that available to me. But more importantly, it made me remember a previous technique I had, using depth to get a shade.

At that time, I wrote this off for the time being because the results weren’t accurate enough, but I did remember it. The depth pass made me consider that I could use it as part of the solution, if not the whole thing itself. The most problematic details from fresnel were at the front. So, I could use the model’s depth as a mix factor to get the nice smoothness it has at the front, and keep the overall shape detail the fresnel captures.

Doing it that way, I was able to get a much more pleasing result.

The green one; the others represent steps in the process.

This new shader uses the depth to get smoothness at the front, and be controllable, while keeping most of the silhouette. I also made some more node groups to add a false paper texture according to how coloured parts are, small variation, and again, noise at the edges. I also added more features to control the shading, like the ability to set the minimum amount of colour an area must always have, for tricky areas or ones needing constant detail. Although, this depth+fresnel method does need tweaking when the view is changed, and relies on the object’s origin being its centre of mas. That does give it some flexibility, too, though, if I wanted to move that to centre a specific point instead of the actual centre.

I’m quite happy with how this looks, for once. It’s not perfect, but it is a step up from previous attempts. There is something else I did differently here, too. The transparency. Ideally, I’d like to use Blender’s hashed transparency. It provides most natural looking results, and looks smooth.

However, I find it’s not as responsive in the viewport. It also takes longer to be able to see what’s going on, since it’s grainy for a moment, and is harder to judge. The difference isn’t too bad, but I’m on a laptop; I’d like to buy all the performance I can get. Especially since it’ll likely multiply significantly in performance loss in a scene with many objects.

So, I considered using blended transparency. It’s quick, it’s smooth. But, it doesn’t seem to work quite right. I have to make sure to set backfaces to be rendered, and then it doesn’t show correctly. Switching that off, however, seems to work.

But it hadn’t worked quite right for me before, and I want to keep performance as good as I can get it, so I thought about the simplest, which is alpha clipping. For that to work, it needs a game type transparency, called Screen Door Transparency. It’s not really transparent, but by hiding more pixels, you give the illusion of it.

Blender doesn’t have anything like that by default, so I made a group node myself, that does it in six steps, converting a Value input to work with it. I think the results are decent, and it’s more responsive in the viewport than hashed. It’s more digital and 3d, though; I may end up using Blend, if it works sufficiently. Disappointingly, the performance at render time compared to hashed is about the same, which is strange. It would be frustrating if Blend works just fine after all the time I spent working out how to do that. Although I do think this way gives distinction to it.

The overall size of the node set is massive, though. I could put it in a group by replacing the colour ramps with Remap Value nodes I made, but the control would be linear and insufficient.

In any case, I’m fairly satisfied with it for now. I can choose the canvas colour, main colour and shadow colour, control the falloff from the centre to emulate watercolour spreading, control transparency to emulate it being thinner the further from that point it is, and control the way the HSV is modified by the fake paper texture.

Next, I want to add more. I want to be able to add fake brush strokes by using a texture as a vector input, which I believe I know how to do but just never did yet; similar way to how a normal map works. I also want the colours to be dynamic, responding to coloured light. That I know how to do already, but need to check more how it would appear in Eevee.

Those can wait, though. My next task is to apply it to a model and make some art. But for now, I’m going to go to bed because my brain is slowly packing it in. It is 1AM, after all.

Shadow Catcher V1

Recently, I made a shadow catcher in Eevee.

An advantage Blender Internal had, and that a particular modified Blender build has with Cycles, was Light Groups. A light group allowed you to cause a light to only affect certain objects. It was convenient for NPR purposes. The other useful, and sorely missed feature Internal had, was the ability to define, in the material, whether it took shadows or not. That’s incredibly useful, because it means you can use it to isolate shadows, not just regular shading, and define which areas you want to take shadows or not.

For example, anime character’s eye and brow, if shaded normally, would likely have a strong and ugly shadow cast by polygon hair, which doesn’t look good. You could, if you could isolate it that way, mask it out. That’s not possible by default in Eevee or Cycles.

But, using Eevee, I’ve found a way. I wanted to be able to isolate the shading from any given light. Especially since it was one of the techniques used in Guilty Gear Xrd to achieve such convincing false 2d.

They used a light on each individual character to give them specific control, rather than scene only lighting. I’d like to explore scene-based shading, lights in the world, etc, but it’s good to have this as an option. That made me consider isolating lighting results again. I did some research, and came across this Polycount thread.

I already knew shading could be faked by using the dot product of a direction and the surface normal, but that can’t capture shadows. I’d thought beforehand, I could use the RGB node to get the difference between fake shading and real shading if I could just make them match; what I lacked was the ability to work out what direction the light is coming from in order to do that, and that’s what that thread taught me.

By using this technique, I was able to get the light direction. Then, by simply using a Shader to RGB node with a diffuse shader and subtracting the false shading from it, I was able to isolate the shadows.

With this, I can catch shadows and then use whatever technique I’d like to mask them out.

The advantages of this techniques are:

Able to identify shadows.
By using multiple lights, you can identify specific shadow types, like using two to distinguish contact shadows and shadows.
Can use multiple to distinguish hard and soft lights, then mix between them to preference.

The disadvantages:

It’s dependent on drivers. You have to manually select the light, and set up drivers for its location.
If the light is hidden, I’ve discovered the driver will break, meaning you have to reassign the object.
If you move it into a different scene and don’t bring the light, you’ll have to replace it and redo the drivers again.
It’s inconvenient to have to use multiple lights to get multiple types of shadow.
Using multiple lights makes the whole scene brighter. Objects not using the shadow catcher can’t account for this, and may be undesirably bright.
It’s inconvenient if the lights have to be changed significantly, like the number.
I believe it only works correctly with directional lights.

Having to set it up in such a time consuming way is the big problem for me. It’s not as if there’s just one box to select it in. It’s inconvenient, and especially when it breaks just from the light being hidden. That might be fine if we use the Guilty Gear Xrd method and only use a specific light or set of lights for it, but if we want to change things a lot, it’ll be a pain.

So, I’ve been doing some more experimenting, trying to recover that information somehow using the Texture Coordinates node, which can select the object. If I can, it will be much easier and quicker to modify. I’ve only had mixed results so far, and not worth showing yet. Still, I’ve learned from this, so it’s not wasted. I’m hoping to get a convenient way working of shadow catching this way. The ability to modify how it’s shaded and shadowed, in real time, not compositing, would be extremely useful.

Frustration

Recently, I’ve been frustrated with my shaders.

None of them look the way I want them to. I’d like to achieve a painterly appearance, or watercolour, but they’re quite elaborate. I sometimes think it would be better if I could settle for cel shading, but I just can’t accept that. Or rather, I want to be able to do more than just shade everything so simply. I’m not one of those people who’s interested in NPR primarily, or purely, to replicate the style of anime. It’s not that I look down on that style; it’s just not my preference.

So, I’ve been doing more research, as usual. The big problem with using Fresnel or Layer Weight to make the rough shape of the object, or an outline, is that it can give unwanted lines, and does. I dislike having to manually correct these, so I’ve been trying to think of ways to fix this. I’ve experimented the day or so with blurring the normals of the model to remove details.

How did I do this? Well, I’m not blurring the normals, exactly. I came across this video a while back. One of the features demonstrated is creating a false blur by distorting a texture.

23:58 for the blur technique.

I’d hoped I could apply the same technique straight to the normals, but it didn’t work. Instead, I baked the object’s original normals out to an image, and used that method to distort it.

The result was effective, actually. I haven’t removed anything from the default Suzanne head, but by blurring it and using it as the Normal input, it’s removed the detail. It also darkened it for some reason, which I was able to compensate for by remapping the value by an equal amount to the distortion. I’d hoped to use this to get a blurred version of Fresnel I could use to make a better silhouette. However, when I tried it..

It only works from the front. I really don’t know why this is. My best guess is that it’s something to do with the normals’ texture. The colours don’t appear the same way when I look at them as the original normals do, so perhaps it’s a reason why it doesn’t display right. However, when I used it as the normals without blurring, it displayed correctly.

It’s quite frustrating. But at least I learned I can blur out detail with this method. That’ll likely come in useful.

The advantage of this method is that the original normals are a texture that isn’t being permanently modified, so it can be controlled as I please.

The disadvantage is, it’s grainy. I haven’t found a way to smooth it yet, though, for what I’m doing, a bit of grain may help it appear natural. Also, as it’s just using the original normals from a texture, it can’t be modified the way the real normals of the geometry can.

On another note, I came across an interesting method for modifying shading.

This technique comes up ~ 1 hour, 15 minutes in.

This one uses an image texture to paint on to control shading. The trick is to create a UV slot on the objects you want to modify, and then use the UV Project modifier to project UVs, like the Project From View UV mapping option, but in real time. Then, have an image texture using that UV to control the shading by adding or subtracting. As they’ll all reference the same image, you don’t need to select individual objects to control, and can modify everything in the shot at once.

The advantage of this method is that you can control them all at once as long as they’re referencing the texture. Additionally, if you make it the same resolution as your render – and you probably should – it shouldn’t appear pixelated. And since it’s an image, you could also edit it externally if you really needed to or wanted to. I find this method far more responsive as I paint than modifying vertex paint, and it can be far more precise than vertex paint would be on most sensible meshes. It’s extremely effective for cel shading.

The disadvantages are several. For one, it’s an image, so it’ll bloat your file size, potentially a lot if you’re rendering at high resolution. Also, because it’s an image, you can’t animate it the same way you could the vertex paint method; you’d have to use an image sequence, increasing the file size further and manually repainting each frame. Although, if replicating anime, that may be best, as it’s essentially what colouring in anime is like, and leaves room for human errors that make it appear more genuine. Lastly, if you change anything in the shot, you’ll have to repaint. Depending on the change, this can be small or large; if you move the camera, you’d have to redo the lot, or, if you were just panning, perhaps you could compensate using the Mapping Node or translation method. That makes it tricky if you want to change the camera angle or re-pose a character or such, though if you modify the shading last, it shouldn’t be as much of an issue. It’s also more difficult to make look natural on something with smooth shading.

Overall, I find there are advantages and disadvantages to this method, but it’s certainly something I’ll keep in mind. I might well use it next time I’m doing a render; I prefer the flexibility of vertex paint, being able to modify things after the fact and the shading changes not be rendered all worthless…But the speed and precision of this texture method isn’t something I can ignore. Custom shading is very likely the key to convincing 2d, or at least an essential.

Anyway, that’s what my research and testing has turned up for the moment. I want to experiment more. Reading a reference from an old Siggraph paper, Painting With Polygons, Video Watercolorization using Bidirectional Texture Advection gave me some ideas; that’s actually where my idea to blur the normals came from. It also seems several steps of their process are ones I’d used in my earlier compositing technique, though applied differently. I’ll try experimenting more and see what I can turn up.

Shader Experiments and Improvement Attempts

My previous renders highlighted a problem I’d not thought of while modeling them.

The soft edges look fine against only the white background. They appear soft and match it nicely, giving at least part of the look I’m going for. But, it doesn’t work with environments, or anything behind the object that doesn’t match the colour of the “canvas”, the white colour that’s being mixed with the object colour.

So, to fix that, I’ve been experimenting with replacing that white edge with transparency.

I tried using the alpha blending mode Hashed first. But I had that the problem of the interior’s faces being visible. That makes a more even silhouette, except that the lines of the front highlight the hole, making it appear more 3D. Hashed is also much noisier; even when I rendered with high samples it appeared grainy.

Note that this image is from behind, not flipped.

I also found the transparency worked too well this way; you could see through the model from one side and out the other. Looking at it from behind, I could see the eye that should only be visible from the front. Thinking of it as something to imitate 2d, this is very wrong; a silhouette in 2d is just a flat shape without depth. You should only be able to see the side facing the camera. I tried to correct this by making the backfaces also transparent, but this lead to some undesirable behaviour, too, like the empty eye socket rendering as a hole. Although, to have it both ways there is impossible, and ideally I wouldn’t have any backfaces showing to the camera to begin with.

In any case, Hashed turned out to be insufficient, so I tried Blend, but it has the same problem. It’s less noisy, at least. I suspect my colour ramp controlling the Alpha is at fault for being too weak, in part. But I need some method to get rid of the faces on the other side of the object; to hide what shouldn’t be seen.

After that, I experimented more with how I was making my silhouette, for a change of pace.

I have another method besides the Fresnel to product the rough shape of an object, using the Normals and Location/Position. I quite like the way this one looks. It’s soft and easy to work with. It’s not as specific as Fresnel or Layer Weight on the edges, which makes it easier to make it look 2d. But that’s also a disadvantage; I have do more to make it match the basic shape of some objects. The rough person-shape I made on the right, for example, has much less colour on the limbs. I also found this wasn’t compatible with my current method of texturing the edges, so they’re too smooth and perfectly neat.

It subtracts the Location from the Position so that when it’s moved in the world the values won’t change. The Scale vector is plugged into Normalise from a previous setup I forgot to disconnect. The cross product with the incoming is what allows it to change with the view.

In the end, I concluded that that doesn’t work for me at the moment. It’s too incompatible with previous node groups I’ve made, and I’d like to be able to reuse them as much as possible. It’s also too much work to make that into a generally accurate silhouette. The trouble with the Fresnel method, on the other hand, is that it starts out too accurate and has to be manually modified.

My current stage of the shader, part-way through modification.

I could use a simpler mesh and the Data Transfer modifier with weights to give a simpler silhouette, but it increases the amount of work to make a character, or other object a default Fresnel edge won’t work well on. I’ll continue to look for methods to improve it.

Dragon Knights And Perfectionism

I’ve been having quite a problem making things for a while.

I like to sculpt, and work on new NPR techniques, and I want to show the ships and characters I like well, but in the end, I haven’t ended up posting anything in a long time. I ended up stuck in a crappy perfectionist mindset, always revising or redoing things, or procrastinating from them because I was afraid of getting it wrong, or that it wouldn’t be good enough. But because of that, I didn’t do anything. I don’t have anything finished to show for my sculpting and shaders. But I was encouraged by my friend recently to try and dial it back, and do simpler things and work my way up, and to try and work on my perfectionism, so I’m trying.

I’ve started with chibis. They’re a lot simpler to model and rig than realistically styled characters. I’ve also been finding they’re easier to shade, to an extent; it might be because they don’t have as many bulges and creases and things that can go wrong and look incorrect. Still, I wasn’t able to make them look the way I wanted them to yet. My shaders need more work.

But, I made something. It feels good to have done that, at least.

Millie and I were talking about our Mirror! AU, in which traits of Yugioh characters are reversed, resulting in very different relationships. Jounouchi and Kaiba (we call him Seto in that au, since he prefers to be friendly and warm) are friends, and we had a headcanon that Jounouchi and Seto game together. Millie also made a cute comic about it. It inspired me to imagine them playing an MMO or something together as Dragon Knights, similar to the Dragoons from Final Fantasy.

Since their signature cards in Yugioh are the Blue Eyes White Dragon (BEWD) and Red Eyes Black Dragon (REBD) I designed them to be reminiscent of those, and similar to the Monster World figures from the manga and Toei anime’s Monster World arc.

My design for Dragon Knights Seto and Jounouchi. I’m not a skilled 2d artist.

Then I modeled and rendered them. I’m not very happy with the render, but it’s something; I’ll have to play with their shaders some more, and learn to make environments. Even a very simple one, like the grass they’re standing in was unfamiliar to me. I’ve spent so long doing characters and being a perfectionist about them, I’ve neglected other areas.

I’m not very satisfied overall with it, but I’m glad I finally did something. I’m grateful to Millie for pushing me to do something like this. I’ll also make a post for myself about what’s currently going right and wrong with the shaders, and what I want to do to improve them. I’ll try and build momentum and make more things, too.

More 3DCoat Tests

I’ve been using 3DCoat quite a bit since I started testing it. I don’t think I can comfortably go back to using Blender’s sculpting much having used it; it just lacks a lot of the features 3DCoat has by default, or, it has them, but they’re awkward to set up, and, at least on my device, the performance is far better in 3DCoat than the equivalent polycount in Blender, and its painting and retopology tools are very convenient.

But, I’m still not entirely solid on how I should be using it. What level of detail to use, how I should block out limbs. The voxel sculpting is convenient for making a base mesh shape by putting together some primitives, but I don’t like that method. It doesn’t feel good for me, and I find it awkward, especially in the absence or convenient transform hotkeys like Blender, where I just have to press one button and move my mouse. Or it could have those, and I just haven’t discovered them yet. I wonder what the best resolutions are to work at when doing various things, such as blocking out, starting to get more detailed, etc. I’d prefer to be able to sculpt without thinking about that, but if I went to a high resolution straight away, I’d end up with something lumpy. Although, that also depends on the brush size.

For starters, I’ve been doing more faces.

I’m not very good at faces. That was one I did a few weeks ago, and it’s quite poor. I didn’t reference, since I was just doing it for practice with the software, mostly, but it’s still bad. It also looks too much like the same face I do if I do a head without thinking about it. I’ve been wanting to gain variety, so I tried doing different heads, too.

I quite liked how that one came out. I like wider noses like that better than sharp and narrow ones. I still need to tackle eyes more, though. They’re something I’m awful at. I should practice more on those, too. I think my current experiments with different detail levels won’t apply to eyes, though, since they’re likely to be detailed just by being the way they are.

I also experimented with some torsos. I didn’t reference, so they’re also flawed, but the main point was learning what works better. This first one, I tried a resolution of 2.00, but using smaller strokes. It was a bad one, I think; I found it became lumpy easily even after smoothing, and didn’t really benefit from it. It was slightly clearer crap.

The next one, I tried a similar level of detail, but using broader brushes. I haven’t tried smoothing this out yet, but I feel better about it; I was able to build up the general shapes more easily, while still being fairly clear. I used the General Clay brush to just build this up and carve bits out, unlike my usual method that uses the General Clay brush to make the general shape, then the Flatten brush to make it into the shape I want.

I still have a few days left on my 3DCoat trial that I want to make the most of. And apply this to proper models.

Testing 3DCoat

Recently, I’ve been playing with the trial version of 3DCoat. A problem I’ve been having with Blender that’s been frustrating me is the sculpting performance. When watching timelapses of other people sculpting, I noticed that often, even with single bits like an arm or a leg, they could easily go into hundreds of thousands of faces and above, especially when smoothing, adding creases, etc. In Blender, I can only get to about 100,000 triangles before it starts to slow down slightly, particularly when trying to undo. At a minimum, with the detail level I use, I need around 200,000 – 300,000 for a full body, but the performance really starts to become slow and problematic. It’s just totally insufficient for me. I don’t tend to make extremely high detail meshes, but I’d like the option, and better performance in general, so I started looking at other digital sculpting programs. I’m ignoring ZBrush; though it’s considered the industry standard as I understand it when it comes to digital sculpting, I’m a hobbyist, and don’t need industry standard….And I certainly don’t need industry standard price. There’s no way I can afford it.

I looked at Mudbox first, as it’s immediately the cheapest one that I know of. It’s about £12 a month, last time I checked, which seems to be the most affordable, at least per payment. But, when I tried out its recently added dynamic tesselation, it wasn’t very covnenient or intuitive, and didn’t give me all the options I’d like, and the options it did have weren’t very well explained, though the performance seemed to be alright. But it also didn’t give me as many hotkeys to change as I wanted, which I hate, as I’m really used to using hotkeys for almost everything in Blender. I also had to pay for one month’s subscription to get full access; the demo version didn’t include the dynamic tesselation, which was the thing I was really interested in as I hate to have to make a base mesh first. I suspect it was deliberate, to extract more money from people.

What I’ve been trying now is 3DCoat, which I’m liking. The workflow is different than I’m used to, but I can see the logic in it, and it gives me many more hotkeys to setup. I was easily able to replace most of the default ones with what I’m used to, and, with a bit of practice, familiarise myself with the different ways some of the brushes, such as the Snake Hook work than what I’m used to, and which ones may be the best options to replace the brushes and settings I’m used to using. I’m enjoying it so far, and the performance, as expected, blows Blender out of the water. I’ve found it fairly easy to get used to, though some things don’t work as I’d expect them to.

So, I did a few test head sculpts to practice and familiarise myself with it and work out how I should work in it. I didn’t reference with these; it was more like doodling than any serious attempt to make something. I need to study more faces and gain variety, though. Especially people from various ethnic origins and ages.

I’m going to spend more time practicing with 3DCoat and see what else I can do with it. But I have a good feeling about it. And I definitely can’t go back to Blender’s sculpting the same way after this. I may well end up buying it.

NPR Shader Experiments 01

I’ve been experimenting for the last few days with more shaders, using Eevee in Blender 2.8. I started trying to replicate my previous compositing effect, then ended up doing something a bit different.

As I can’t use Dilate/Erode with shaders to get a soft edge, I used a Layer Weight node. I wanted to use the Bump node to modify the normals to gain control over how that turned out, using vertex paint to define which normals are bumped, but it ended up displaying incorrectly, so my current solution is to use vertex paint to add or remove colour from the initial appearance, with a second set of vertex paint to modify the shading. I like this method better than using modified normals; I can directly control it, rather than just having to change normals and hope the outcome is correct, or having to change them frequently depending on the lighting or model angle. However, unlike Blender Internal, I can’t currently get a difference in the shader between its own shadows and shadows cast on it by other things, so if I modified the shading for being ugly in an area shadowed by a second object and itself, it would erase the other’s shadow, looking incorrect.

But, if the AnimAll addon is upgraded to work with Blender 2.8, the vertex data will make it animatable, meaning I could easily fix shading every frame. Currently, that addon isn’t compatible with Blender 2.8, though.

I also used a small amount of bump mapping to add a little bit of inconsistency and texture to the models; complete perfection in the shading is a big tell of 3D. I also used a noise texture on the edges of the Layer Weight’s result, to make the imperfect.

I have to work more on this; I think it’s better to combine some of the vertex paint effects, for one. If I’m changing the shading, it should also effect the edge. As it stands, shding that’s removed doesn’t overwrite the original Layer Weight result, so it can cause areas to exist that look too dark next to modified shading. Also, I need to see how it works with different colours. Past experiments have shown dark colours, high saturation, low saturation, behave differently, so I may need to modify how strongly they’re visible according to those to make it look good. I’m making progress, though. I’ll also experiment with applying this to proper models instead of test models, since the purpose is to be actually usable.