<em>Inspired 3D:</em> Subdivision Modeling Techniques

[1]

All images from Inspired 3D Modeling and Texture Mapping by Tom Capizzi, series edited by Kyle Clark and Michael Ford. Reprinted with permission. [1]

This is the third in a number of adaptations from the new Inspired series published by Premier Press. Comprised of four titles and edited by Kyle Clark and Michael Ford, these books are designed to provide animators and curious moviegoers with tips and tricks from Hollywood veterans. The following is excerpted from Modeling & Texture Mapping.

Subdivision modeling techniques are used to take a low-resolution polygonal object and increase the resolution using a smoothing algorithm to create a high-resolution model. Several methods work quite well to accomplish this task.

Polygon smoothing is conceptually the simplest type of subdivision modeling. The original polygonal model (Figure 1) is defined as the low-resolution cage, and the higher-resolution geometry is created directly from it (Figure 2). You can use subdivision steps to determine the final resolution of the resultant model. As a rule, the resolution should begin with one single subdivision and increase from there based on the need of the model. The entire model can be subdivided, or selected faces can be subdivided.

[2]

[Figures 1, 2] Polygon smoothing is a predictable and easy way to create a high-resolution model from a low-resolution model. [2]

The advantages of using this technique are as follows:

As long as the history is maintained, the model can be returned to the original state by selecting the smooth node and dialing the subdivision number back to 0. Editing the high-resolution mesh, however, will cause unpredictable results if the subdivisions are set back to 0.
The resultant geometry type is polygons. Maya has few problems dealing with polygons and will behave in a stable way when using them.
UVs are maintained in a predictable manner.

Some disadvantages include the following:

Some artifacts can appear at the edges of models because the polygonal density is not high enough. This artifact, called nickeling, can be fixed by increasing the density of the polygons at the edge of the model, or by increasing the number of subdivisions in the smoothing operation.
The ability to interactively work on a low-resolution polygonal model while previewing a high-resolution view of the smoothed model is not available in Maya. This can be done, however, using a plug-in called connectPolyShape, which is available at www.highend3D.com [3]. This plug-in can change the way a modeler works and is definitely worth checking out.

Subdivision surfaces use an internal interpretation of the polygonal mesh into another entity type. This entity type behaves similarly to NURBS surfaces. The easiest way to understand this process is to look at Pixar’s RenderMan, which creates subdivision surfaces.

Within a low-resolution cage are quadrangles and triangles. In RenderMan, these entity types are treated differently. A quadrangle is assigned a NURBS surface. Every quadrangle in a polygonal mesh has an infinitely smooth surface that is tangent to the adjacent surfaces. At render time, these surfaces are tessellated adaptively at a pixel level. This unique tessellation method allows for unbelievable detail when rendering displacement maps on relatively simple surfaces.

Triangles, however, are not defined as NURBS surfaces. They are defined as subdivided triangles, in a similar way that smoothed polygons behave. The ability to tessellate adaptively is reduced.

Maya behaves differently than RenderMan. The geometry can still displace better than any smoothed polygon model, and the areas where quads transition into triangles are treated differently.

These points can be seen as advantages that subdivision surfaces in Maya have over smoothed polygonal models.

Other advantages include the following:

Interactive editing of low-resolution polygonal cage while previewing high-resolution geometry without using plug-ins.
Many tools that allow quick editing, mirroring and conversion of subdivision surfaces.

In short, Maya has developed many tools that make subdivision surfaces look attractive. But it must be noted that this entity type is notoriously unstable. Before using this entity type on a production, test it carefully and often. Results attributed to using subdivision surfaces include these:

Loss of UV information, especially across mirrored axes.
Maya has an invisible node called the shape node associated with every piece of geometry in the scene. Maya uses these shape nodes in the dependency graph for many important functions. Using subdivision surfaces can cause geometry shape nodes to simply disappear. Digging through the hypergraph can get the geometry back, but only after a heart attack or two.

Detail
Detailing in polygonal modeling has to be done in combination with a way to preview the results, which is why the smoothing discussion was introduced before the discussion on creating detail. If the resultant model is going to be smoothed using subdivision modeling techniques, then the results of this additional process should be checked whenever a significant amount of work is to be done. When the lips are detailed, check them, when the ear is detailed, check it and so forth.

Detailing usually requires the model to be split along the areas where the model has a topological change. For example, the edge of the lip is not exactly a hard edge. But if the edge of the lip is compared to the side of the cheek, it is significantly sharper.

Creating detail in regions like this requires the process of adding additional rows of polygons along these areas. To create the ridge at the edge of the lip, a row of polygons is created at the edge of the lip; when this single row is subdivided, it becomes two or more rows, adding more definition.

When applying additional rows to create detail, it is important to understand how these rows will affect the final model. Some simple rules can come in handy when these conditions arise. In the examples in Figures 3 through 10, different examples of polygonal smoothing are shown.


[Figures 3, 4] A model with no rows of controlling polygons.	[Figures 5, 6] A single row of controlling polygons.

[Figures 7, 8] Additional geometry added at the corner.	[Figures 9, 10] A model with two rows of controlling polygons.

Sharp corners will smooth out if there are no additional rows of polygons inserted (Figures 3 and 4). Additional rows of polygons at the edges and corners help control the way the geometry is smoothed. These additional rows of polygons are used to create areas of detail in the final model.
In Figure 5, the shading artifact that blends through the single row all the way to the corner is called flashing. A single row of polygons will not stop flashing along the face of the square. In Figure 6, the corner where the rows come together was not controlled by adding an additional polygon, so the corner was smoothed unpredictably. A single row of polygons works better than no rows at all, but will not provide adequate control for detailed areas.
In Figure 7, additional polygons were added at the corner in the image below left. This allowed the smoothing operation to behave more predictably in Figure 8.
In Figure 9, additional rows of polygons were added along the edges. Notice how the highlights on the edges are confined to the two rows. In order to control flashing, a large face on a polygonal model that transitions into a smaller face must be separated by two rows of polygons. Figure 10 shows how the additional rows give the smoothing operation more control.

[4]

[Figures 11, 12] The low-resolution ear (left) and the high-resolution ear. [4]

The areas of the face that normally require additional work are the ears (Figures 11 and 12), eyes (Figures 13 and 14), nose and mouth (Figures 15 and 16). The details range from major reconstruction to the simple addition of a polygon row to sharpen the area just a small amount.

[5]

[Figures 13, 14] The low-resolution eye (left) and the high-resolution eye. [5]

[6]

[Figures 15, 16] The low-resolution nose and mouth (left), and the high-resolution nose and mouth. [6]

Cleanup
Once the polygons have been split, sculpted, merged, deleted and manipulated into the model that is going to be smoothed, certain cleanup tools should be used. In truth, these tools must be used every time the model is going to be previewed using the subdivision method required for the model. These first tools that should be used are merge vertices and merge multiple edges. These will simplify the unseen entities that may be creating problems.

To do a final check on the model, you can use the polygon cleanup tool. You must use this tool carefully. Be sure that the model is inspected carefully before the results of this tool are accepted; this tool can cause major problems to otherwise usable models.

Hair
The original sketch had a baseball cap on the head of the character. This was an attempt to avoid what became a difficult process of making many layers of NURBS surfaces into hair. Many computer-generated characters use layers of NURBS surfaces to create hair. This character was supposed to be a young guy who did not pay careful attention to hair care, so the hairstyle would have to be loose.

The real story here is the difference between the hair that was originally proposed and what finally appeared on the character.

Figure 17 and 18 show some of the progression from long hair to the relatively clean-cut look.

[7]

[Figures 17, 18] The first hair proposal (left) and the final hair. [7]

Eyes
The eyes were built from three NURBS spheres nested inside each other: A clear outer layer (the cornea). This layer is simply a clear reflective ball that surrounds the rest of the eye.

A colored interior layer (the iris). This layer has a recessed, or concave, area around the color of the iris that reacts to light. When light is directed above the eye, the iris will have additional reflection that occurs beneath the pupil. This anatomy is physiologically incorrect, but this lighting has become an accepted way that eyes appear in computer-generated characters. The characters in the Pixar films and the PDI films have their eyes constructed in a similar manner. The opening for the pupil is simply a hole that has a cluster that controls the diameter of the hole. This allows for animation of the size of the pupil.

A black inner layer (the pupil). This layer is adapted to fit the iris. The shader on this layer is a black surface shader that emits no light whatsoever.

[8]

[Figure 19] A close-up of the eye reveals some of the way the eye was built. [8]

Teeth, Gums and Tongue
These were simple models. The teeth are simple NURBS spheres that have been flattened to resemble teeth, and the gums are surfaces that have been sculpted to accept the simple teeth.

The tongue is a half sphere that has been modified to resemble a tongue. All of these items are NURBS construction because this type of model is relatively easy to create, map, and animate. Even though these parts are important, the focus of this chapter does not need to cover this material.

Facial Animation and Blend Shapes
In the production of an animated character, the character can have the face animated in two basic ways. One way is to have animation setup control the face using various setup techniques. This requires expertise on the part of the setup technical directors. In a large production facility, the efficiency of scale can make difficult jobs like this commonplace. Many characters have already been set up that can be taken apart and reused. In a small production, the process of creating facial controls can be time consuming. This is especially true because no other similar characters can have their controls “recycled” for the new character.

On this production, despite my best efforts to avoid this, the production decided early on that the facial animation would be controlled using blend shapes. Blend shapes are 3D morph targets that have the exact topology as the face they are controlling.

Full Face Shapes versus Local Face Shapes
In the creation of face shapes for an animated character, there are two basic schools of thought regarding the way the face should be animated using blend shapes. One method is to use the entire face as a specific target. If the character is going to frown, then the entire face is sculpted into a frown shape. The eyebrows are sculpted into a furrowed appearance, and the entire mouth is sculpted into a real scowl.

Face targets like this are normally sculpted to the maximum range, which can be used as the production dictates. If the shape is a 100% frown, then the production can use this shape in increments of 20%, 40% and so on, to create frowns of less intensity.

Local face shapes are broken down into specific regions. These into left, right and center regions. If there was a smile to be modeled, the blend shape would be a left smile, a right smile and a center smile. This gives the animator a lot of control as to what part of the face will be affected by the blend shape.

For this project, the local blend shapes were a necessity. But because I know that the vast majority of animation time using blend shapes was spent trying to get multiple blend shapes to animate as a single channel, I also added some blend shapes that took up an entire region. I modeled a left smile and a right smile, and I also modeled an overall smile as well. I modeled a left furrowed brow and a right furrowed brow, and I modeled an overall furrowed brow.

The jaw was to be animated using a skeletal setup. The jaw position is not animated using blend shapes. The modeling of the blend shapes had to be coordinated with the animation of the jaw. The figures below show the jaw in various stages of being opened or shut along with the blend shapes being shown. This helped visualize how the blend shapes would behave during the modeling process.

[9]

[Figures 20-25] Blend shapes that affect larger, less localized areas of the face. [9]

I was careful to try to keep the mouth blend shapes from affecting any other regions, and the eye blend shapes localized to the eye region. In this way, these particular shapes were hybrid global/local blend shapes. They affected an entire area, but only the area intended to animate, not the entire face.

When I was creating the localized blend shapes, the areas that were affected stayed on one side of the face. There were many shapes not shown here that broke each part of the face into even smaller regions. These regions were isolated to areas like one eye (Figures 26 and 27), one eyebrow (Figure 29) or one corner of the mouth (Figure 30).

[10]

[Figures 26-31] Localized blend shapes. [10]

Most blend shapes are modeled about 20% past the most extreme point where the animator is expected to use them. This allows for more elastic animation and gives greater flexibility when combining blend shapes.

Shapes and Phonemes
The decision of which shapes to build came from two primary sources. First, Rick Grandy, the technical editor for this book, came up with a preliminary list, and then the animator, Kyle Clark, came up with some items that he needed to get this project done. Overall there were 66 targets built for this animation, and there will probably be some more that need to be built as more animation is done. Some models used for production have more than 200 blend shapes modeled. The face shapes were broken down by region:

Eyebrows, left, center and right. The shapes created for this region allowed the eyebrows to animate up and down, slide inward toward the center or out away from the center, and bow in the middle upward and downward. Shapes for the brow animate the eyebrows into even smaller regions: the left, right and center for the left brow, and the left, right and center for the right brow.
Eyelids, left and right. The shapes made for the eyelids allowed for the eye to close, by pulling the upper lid down, and allowed the eye to squint, by pulling the upper and lower lids to meet in the center.
Face (for broad shapes), left and right. The face groups had shapes that animated the cheeks up and down, moved the cheeks in and out and puffed the cheeks out and sucked them in. There was also some cheek deformation on broad mouth shapes, such as the dread, sneer, smirk and grin.
Mouth (the largest group), left, center and right. These shapes created simple as well as complex mouth movement. The simple movement includes moving the each lip up, down, curl in, curl out, corner up, corner down, corner side movement inward, corner side movement outward and shapes that smoothed out the corners of the mouth.
The complex shapes required the modeling of the frown, smile, furrowing, puckering, pouting, yawning and kissing.
Overall face shapes, localized by region. These shapes simply used large areas of the face to accomplish a single task. This kind of approach is preferable when there is a specific target that the animator may want to hit with a single blend shape.

These shapes include mouth smirk, mouth sneer, mouth dread, mouth wince, eye furrow, eye squint and mouth smile.

The mouth regions were extended to include the cheeks, and the eye regions were extended to include the forehead and eyebrows.

Phonemes are face shapes directly related to speech. Different theories exist as to which phonemes are required for animation of speech. Thirteen accepted shapes are recognized as vismemes, which are used in the creation of English speech.

These shapes are as follows:

Closed mouth: P in pie, B in book, M in mother.
Pursed lips: W in wicked, OO in root.
Rounded lips, corners of the mouth slightly puckered: R at the beginning of a word, OO in book.
Lower lip drawn to upper teeth: V in victory, F in French.
Tongue between teeth with gaps on the side of the tongue: TH in think.
Tongue behind teeth with gaps on each side of tongue: L in look.
Relaxed mouth, mostly closed teeth, tongue visible behind the teeth: D in dog, T in tag, Z in zebra, S in sit, R in car, N in nothing.
Slightly open mouth, mostly closed teeth, corners of the lips slightly tightened: CHI in chime, JI in jive, SH in shy, VI in vision.
Slightly open mouth, mostly closed teeth: Y yawn, G in get, K in kitchen.
Wide mouth, slightly open lips: EA in meat, I in rip.
Neutral mouth, teeth slightly parted, jaw dropped slightly: E in bet, U in but, AI in bait.
Round lips, jaw dropped slightly: OA in toad, O in rope.
Open mouth, jaw dropped: A in math, O in shop.

For the list of shapes that would be modeled for this model, the phonemes were reduced to eight basic shapes:

Wide, slightly open lips: E in evening.
Round lips, jaw slightly open: O in oh, O in toast.
Round lips, corners of the mouth puckered: OO in book.
Closed mouth: P in pie, B in book.
Lower lip drawn to upper teeth: F in fine, V in vase.
Lips pursed: W in work.
Mouth open, tongue visible from inside mouth: T in tank, D in dog.
Relaxed mouth, mostly closed teeth, tongue visible behind the teeth: S in sit.

To learn more about constructing 3D characters and other topics of interest to animators, check out Inspired 3D Modeling and Texture Mapping by Tom Capizzi; series edited by Kyle Clark and Michael Ford: Premier Press, 2002. 272 pages with illustrations. ISBN 1-931841-50-0 (US $59.99) Read [11] more about all four titles in the Inspired series and check back to VFXWorld frequently to read new excerpts.


Author Tom Capizzi (left), series editor Mike Ford (center) and series editor Kyle Clark (right).

Tom Capizzi is a technical director at Rhythm and Hues Studios. He has teaching experience at such respected schools as Center for Creative Studies in Detroit, Academy of Art in San Francisco, and Art Center College of Design in Pasadena. He has been in film production in LA as a modeling and lighting technical director on many feature productions including Dr. Doolittle 2, The Flintstones: Viva Rock Vegas, Stuart Little, Mystery Men, Babe 2: Pig in the City and Mouse Hunt.

Series editor Kyle Clark is a lead animator at Microsoft's Digital Anvil Studios and co-founder of Animation Foundation. He majored in Film, Video and Computer Animation at USC and has since worked on a number of feature, commercial and game projects. He has also taught at various schools including San Francisco Academy of Art College, San Francisco State University, UCLA School of Design and Texas A&M University.

Michael Ford, series editor, is a senior technical animator at Sony Pictures Imageworks and co-founder of Animation Foundation. A graduate of UCLA’s School of Design, he has since worked on numerous feature and commercial projects at ILM, Centropolis FX and Digital Magic. He has lectured at the UCLA School of Design, USC, DeAnza College and San Francisco Academy of Art College.