SUBDIVISION SURFACE MODELLING

What is subdivision surface modelling?

The most flexible and powerful techniques in professional modelling leave you with the most versatile geometry for any task.

This is what we strive to create with subdivision surface modelling. We want the subdivision surface modifier to approximate geometry for us. We don't model curvature, we control it.

We define start and end points for curvature which the subdivision surface modifier uses to perfectly create the hard or soft edges we want to see.

That's the principle. The practice is a little more involved!

Introduction

We have access to incredibly sophisticated hardware and software which makes the creation and rendering of 3d objects a trivial task. Connect some points together and make them into faces. Add some light and tell the software that a point in space you define is a camera. With just these simple steps, mesmerising imagery can be created in minutes.

The software, Blender in our case, gives you hundreds of ways to make things and a lot of freedom to create topology structured in virtually limitless combinations. If the results look good in your render then perhaps that is just fine.

More and more however, the expectation is that the 3D rendered images we create closely follow what we would expect to see in real life. We care about how light flows over and reacts to surfaces. Modern design involves the use of hundreds of interconnecting curved surfaces and edges and we want to be able to control all of that curvature individually.

In Blender, this fluidity over our surfaces is achieved through the use of a "subdivision surface modifier". The incredible cohesion and continuity offered by the algorithm comes at a cost. Many of the modelling methods used in free-form polygonal modelling will break this continuity. The power of the subdivision surface modifier means a new approach and understanding of topology is required.

It is most certainly worth the effort and the new understandings gained will make you a better modeller in every situation. That is what I aim to teach on this website

The Framework Mesh

In subdivision surface modelling we are interested in creating the Framework Mesh for any given object.

Many versions and resolutions of a model will be required during any production and these are all derived from what we call the Framework Mesh.

Multiple resolutions

Models in games (Unity, Unreal, Godot, Cry-engine, Lumberyard, Roblox and more) and for use across various formats (Web, phone, tablet, console, PC, 1080p, 2K, 4k, 16K and beyond) have different requirements to suit the display technology and the usage of the model.

A model exported for use in Unreal or Unity for a PC or console game will very have very different requirements to the model used in Roblox (A current limit of 10,000 faces per model in Roblox).

Unreal's 'Nanite' system means that virtually any density of geometry is allowed so very close up detail can be imported into it and the LOD (Level of Detail) system it creates will decide dynamically what is shown.

Another version of the same game in Unity and Unreal may be destined for a mobile device. Other versions will be required to make cut scenes , posters, advertising materials and promotional printed merchandise.

Many of these output models will contain triangles, n-gons and multi-spoked poles. As long as the silhouette can be maintained and the edge normals transferred believably, these things are all common practice.

Every one of these different versions of the model are derived from the Framework Mesh.

Blender Framework mesh and derived versions

Level of Detail (LOD) from a Framework Mesh

LOD (Level Of Detail at various draw distances) models for games are created to preserve the highest levels of fidelity at various view distances.

These LOD models are all derived from the Framework Mesh using a variety of decimation tools and often some manual removal of geometry to get the best version of the original framework model possible at a new resolution.

A model may spend much of it's time in production as a prop sitting benignly in the background of a scene. and for this a low resolution triangulated version will be made. It may contain hundreds of N-gons. It may have all of the faces which can never be seen by the camera removed. This is usual. The original mesh that the triangulated, n-gon filled prop came from will be an all quad, subdivision surface model - the Framework Mesh.

Deformation

When one of the story telling or creative team, reveals that a missile is about to hit the scene or a superhero is going to start smashing things up or the temperature is going to increase and everything needs to melt or a black hole is going to open and everything will be stretched and squeezed into it or everything needs to turn into semi transparent ice; the models need to be ready for this abuse or they will not be used throughout the entire production.

The continuity team will not allow a gun, a vase or a telephone that was there in one scene to be missing in another scene because the model could not handle the required distortions to it's geometry. Then the Framework Mesh itself will come into play and your models need to be ready to handle any kind of damage and destruction the scene requires.

Remastering

Whenever a game is ported, updated, rebooted, remastered or a sequel is required, new models are again based on this original Framework Mesh.

A new mechanic may be required to have a sword swim around its wielder like a fish would. Collected coins may bend and squash in a new way. Guns may get caught in machinery, bent and squashed. Equipment may need to bend and tear through heavy use. You can never predict the future needs of your models but you can prepare for them by ensuring the Framework Mesh, from which they are derived, has good topology.

When a model needs to be remade to update a project, that process costs time and money and no production wants to tolerate this.

If you were to consider only the current usage and requirements for a model, then you remove all of the required functionality of that model in even the simplest of production pipelines.

Models with n-gons are considered unusable as a Framework Mesh in a production pipeline and any models containing them will be rejected at some point in the production. Triangulation of these quads (which is particularly unreliable with non-planar n-gons) is not enough to retain control over the surface curvature and "Decimation" (the process of reducing the face count in a mesh) is too unpredictable a method of polygon reduction to be considered for Framework Meshes; particularly when UV maps are involved.

UV Unwrapping

When it comes to UV unwrapping, quad based subdivision surfaces allow the most predictable UV maps with the largest islands, least stretching and most predictable placement of seams.

With a good Subdivision surface the marking of loops to control the UV seams is almost automatic and will always give the best results.

Models with triangles make seam selection difficult and models with N-gons can be particularly unpredictable as blender's UV system struggles with Convex polygons, which are most often associated with N-gons.

When a subdivision surface Framework Mesh is created, UV maps can very easily adapt to any given resolution whereas any other topology will cause problems.

Subdivision Framework Meshes

The models we use in professional productions are always suitable candidates to have a subdivision surface modifier applied to them which will increase and decrease the resolution of the model without changing the details created by the modeller. These are the Framework Meshes.

The Framework Mesh will be made from all quads using as few vertices as possible. Performance in a scene is essential when working and we need to know that we can have many hundreds of objects without blender slowing down or locking up completely because we have too many vertices; all the while knowing that the render will be perfect because the subdivision levels can be increased to any level required to render the scene perfectly. Anything else is a compromise which will restrict creative possibilities and slow down production pipelines. This is not acceptable.

When selling, licencing and distributing your models, it is often much cheaper for the production to buy pre-triangulated versions. The Framework Models are considered in the same way that Raw files are considered in photography and are the most valuable assets a modeller has. When employed directly by a studio, it is usual that ownership of the Framework Meshes belongs to that production - not the modeler!

What does blender use?

Our current understanding of geometry and topology in computer graphics is centred largely around the use of “subdivision surfaces”. Mostly variations on a specific algorithm created by Edwin Catmull and Jim Clark in 1978. It is effectively way a way of taking a simple geometric description of an object in the form of a mesh (often called 'the cage'), and increasing its resolution in such a way that smoothing occurs across it’s surface. Many other subdivision techniques and algorithms are available but the industry has settled on using Catmul-Clark so that is what we will concern ourselves with here.

This is achieved in blender by creating a polygonal model and applying a Subdivision surface modifier to it.

There are four basic steps to subdividing a mesh of control-points:

1. Add a new point to each face.
2. Add a new point to each edge.
3. Move the original points to a new averaged position.
4. Connect the new points together.

That is it. As modellers we dont need to worry about how this is achived. It is accepted that this method works reliably and we can base all of our models on topology which suits this process. The rest of this website will try to detail what that topology is.

What is polygonal modelling?

Polygonal modelling is the freedom to move points and extrude polygons any way you want to achieve a hard edged shape depicting an object.

Polygonal modelling has no care for the curvature around edges and corners, and does not attempt to restrict the use of n-gons, triangles or the placement of poles. The resolution of a polygonal model is not considered to be adaptable so they are not used to make a Framework Mesh.

Any apparent curvature in a polygonal model is usually the result of bevelling edges to create a surface which, from a given distance will look acceptably smooth as it should in real life. The resolution of this curvature is fixed at the time of it's creation and cannot be easily changed latter in the process. Bevelling is used in subdivision surface modelling, but not to create curvature! We use it in a very specific way to define the limits of curvature. I will go into it later!

When modelling polygonaly, you have complete freedom to make whatever shape you want, using any arrangement of connected faces you desire. The biggest problem you will come up against in polygonal modelling is Z-fighting (when two faces occupy the same space and the rendering engines do not know which one to display on top).

Polygonal modelling is easy and flexible. However, the results are not analogous with objects in real life. The corners and edges are impossibly sharp and that is the nature of polygonal modelling. This may be exactly what you want and an entire sub genre of modelling concerns itself with low-poly descriptions of objects which can most certainly look incredible.

If you want an object to look as it does in real life, you will find that this modelling style will fail in many areas. You can force the renderers to cheat by recalculating normals (the vector perpendicular to a face, edge or vertex), but you are still going to have impossibly sharp edges or approximated curvature and that does not help us to describe how objects react to light in real life.

Beginning modellers often think that polygonal modelling and subdivision surface modelling are the same thing which results in the creation of restrictive models where essential, descriptive detail needs to be hidden.

Booleans and bevels are used without understanding the problems these things can create. Both of these processes can stop the flow around a model's topology and change the way the geometry reacts to light. The problems Booleans and bevels create, while fixable, are usually left as they are, and tricks with normals and creasing are used to hide the problems. The models are not framework meshes. They are not flexible in a production pipeline and close-up renderings of parts of your model are not possible at all.

Animation, UV unwrapping, warping, soft-body, cloth, transparency, sub-surface scattering, rigging and distortion become unpredictable and many models are either rejected further down the CG pipeline or the end results are unappealing, unbelievable and distracting.

What is subdivision surface modelling?

We want to describe a shape as efficiently as possible so that we can apply a subdivision surface modifier to create the final geometry (the Framework Mesh) which describes our final model.

We want it to consist entirely of quads. We want the quads which make up each local region of geometry to be of as similar a size as we can make them. Each of these local geometries will be logically connected together using boundary zones which define the limits of edge curvature.

Subdivision Surface modelling seems like polygon modelling but you do not have nearly as many freedoms to do whatever you want to create a shape. The advantages Sub-D modelling provides, allows organic curvature over the surface of a mesh which is adaptable, deformable, and not resolution restricted.

The way the subdivision surface algorithms divide your geometry comes with restrictions. Triangles and n-gons (faces with more than four connected edges) create pinching effects. Vertices with anything other than 4 spokes change the way light gathers around curvature. Vertices with more than 5 spokes produce unacceptable wrinkling and pinching of light so we do not allow them to exist. There is a valid argument for the use of 6 spoked vertices under certain circumstances but we will do our best to avoid them anyway until we understand exactly what they do to the logical structure of an object (They do have uses in creating Text and Logos from curves which you can see when we recreate a type face for use in blender.)

The bevel operation needs to be used very carefully as it will attempt to change our underlying geometry and can force our subdivision surfaces to become polygonal again. We do use the bevel operation but not to describe curvature. we use it to define the boundaries of curvature which the subdivision surface modifier can then use.

We understand Boolean operations are largely incompatible with subdivision surfaces and will always try to force our model to be polygonal. We don't use them often but when they are used, many complicated processed are often required to fix the problems they create.

We inset polygons. We don’t extrude them. Instead we extrude loops (or we fix extruded polygons as soon as they have been extruded as they will always place a 5-spoked pole on an internal corner.)

‘Sculpted’ geometry has to be completely rebuilt with all of these restrictions considered.

There are definitely going to be occasions where you cannot achieve all of the ideals in subdivision surface modelling. You have to be able to adapt to any situation. The techniques I will talk about are worth aiming for but they are not hard and fast rules. Consider everything you have learned about modelling from every source and try to find ways to make them all work together. Compromise and communication with everyone in a computer graphics pipeline is essential. As soon as your models are required in a production, you will quickly, and often painfully, learn what is acceptable and what is not.

There are other times when a subdivision solution to a model would be a terrible waste of everyone’s time. A textured cube in the distance will make for a perfectly acceptable building in your shot but anytime you need to see close-up detail on an object then subdivision surface modelling techniques will be required. our job as modellers is to help maintain an illusion. It's all smoke and mirrors as it has always been. Believability in context is the goal.

Again, we will not be using Booleans in subdivision surface modelling to create final geometry. I do use them for other things (making patches to attach to models, creating cages and colliders for fluid dynamics simulation for example) but they can mess things up really quickly in subdivision surface modelling so we will avoid them at all costs. As with everything, don’t discount them completely. You will find uses for them all over blender but, in subdivision surface modelling, they should only be considered when you have exhausted all other avenues!

Hard Surface vs Organic Modelling

You will hear people talk about hard-surface modelling and organic modelling and there are constant debates about what models fall into which category. At the moment, current distinctions are just too vague. We want to be subdivision surface modellers and that is all. Feel free to get involved in that debate; perhaps it would be useful to have some defined rules for which is which. But I find the discussion unhelpful at this stage. There are currently too many different ideas and they all have their disadvantages and merits.

I might propose that the best definition between the two types of modelling is that any model that has (or would believably benefit from) an internal joint is an organic model. Any object that does not - is a hard surface model. The shape of the model does not matter. The organic nature of the object as it is defined in life sciences does not matter. An apple without its stalk would be a hard surface model. An Apple with it's stalk would be an organic model.

As before, you really have no idea what your models might ultimately be used for as the scope of a project grows and changes. Your apple may need to be squashed or crushed. It may need to appear be made of steel or rubber. It could be dented, have worms crawling under it's surface, melt into wax, rot away or have it's skin peeled. If you keep as closely as you can to the guiding principles of subdivision surface modelling then the distinctions become irrelevant. Your model is flexible and that is what productions care about.