Subdivision Surfaces

Nearly every animated CG character you have ever seen — Woody, Elsa, a Na'vi, a talking sponge — is, underneath, a subdivision surface. The modeller never sculpts the millions of tiny polygons you see on screen. Instead they push around a small, blocky control cage of a few hundred faces, and a purely mechanical rule grows the smooth surface from it: split every face into smaller ones, then nudge every corner toward the average of its neighbours, and repeat. Do it a few times and the jagged cage melts into something glassy-smooth; do it forever and it converges to an ideal limit surface.

That single trick — refine, then smooth, then repeat — is the workhorse geometry of the whole animation pipeline. This page builds it from the ground up: the two half-steps, the exact averaging masks Pixar's Catmull–Clark scheme uses on quad meshes, the triangle-mesh cousin Loop, the awkward extraordinary vertices where the surface goes slightly soft, and the creases that let you carve a sharp edge back in on demand.

Two half-steps: refine and smooth

A subdivision scheme is one iteration built from two moves, applied to the whole mesh at once:

Refine (split). Insert new vertices — one per face, one per edge — and reconnect them so every original face is replaced by several smaller ones. The mesh gets denser but keeps roughly the same shape.
Smooth (reposition). Move every vertex — new and old — to a weighted average of nearby vertices. This is where the corners round off. The weights (the masks) are the whole design of the scheme.

Refining alone would just chop the cage into more flat pieces (same faceted look, more triangles). Smoothing alone would collapse the mesh toward its centroid. It is the alternation of the two — split, average, split, average — that pulls a coarse polygon toward a smooth curved surface without shrinking it away. Each pass roughly halves the visible faceting; after three or four passes the eye can no longer see the polygons at all.

Watch it converge

The idea is easiest to see on a 2-D cross-section — a polygon standing in for the surface's silhouette. Below, the faint outline is a deliberately jagged five-point control cage. Press Play (or step forward) to apply the refine-and-smooth rule again and again: each pass cuts every corner and averages, and the jagged star relaxes toward a smooth, star-shaped limit curve. The 3-D surface case is exactly this, run over faces and edges instead of a single loop of points.

Notice the limit curve stays near the cage but touches none of its sharp corners — the cage controls the surface without lying on it. Drag a single cage point and the whole smooth curve nearby glides to follow: that local, predictable control is exactly what makes the cage a joy to animate.

Catmull–Clark: the quad-mesh standard

The dominant scheme for quad (four-sided) meshes is Catmull–Clark, published in 1978 by Edwin Catmull and Jim Clark — the same Ed Catmull who went on to co-found Pixar. One Catmull–Clark pass introduces three kinds of new point, each an average with fixed weights.

Face point F: for each face, the average of that face's corner vertices. F = \frac{1}{k}\sum_{i=1}^{k} v_i
Edge point E: for each edge, the average of its two endpoints and the two adjacent face points. E = \frac{v_1 + v_2 + F_1 + F_2}{4}
Updated vertex v': each original vertex of valence n moves to a blend of the average adjacent face point \bar F, the average adjacent edge midpoint \bar R, and its old position P: v' = \frac{\bar F + 2\bar R + (n-3)\,P}{n}

New quad faces are then formed by connecting each face point to the edge points around it. The result is a mesh with four times as many faces, every one a quadrilateral. Repeat, and the surface converges to a curve that is C^2 (smoothly curved) almost everywhere — the surface modellers actually ship.

Worked example: one Catmull–Clark step on a cube

Start with the humblest quad mesh: a cube. It has F = 6 faces, E = 12 edges, and V = 8 vertices, each of valence 3. Apply one Catmull–Clark pass and count what appears.

6 face points — one at the centre of each square face.
12 edge points — one on each edge, pulled slightly inward toward the two neighbouring face centres.
8 updated vertices — the old corners, each drawn in toward the body of the cube.

Every original square is cut into 4 smaller quads (corner → edge point → face point → other edge point), so the face count goes 6 \to 4 \times 6 = 24. Tallying the rest: there are now 6 + 12 + 8 = 26 vertices, and Euler's formula V - E + F = 2 pins the edges at 26 - E + 24 = 2 \Rightarrow E = 48. After a single step the blocky cube has already become a rounded, barrel-like solid — and it is still made entirely of quads, ready for the next pass. (Its 8 original corners are the only extraordinary vertices; see the warning below.)

Each pass quadruples the face count, so from a cube's 6 faces you get 6 \cdot 4 = 24, then 96, then 384, then 1536… In general a control mesh of F_0 faces has F_0 \cdot 4^{k} faces after k passes — exponential growth, which is why you subdivide only as far as the camera needs and let the renderer do the last levels adaptively.

Loop: the triangle-mesh cousin

Catmull–Clark wants quads. For meshes made of triangles, the standard is Loop subdivision (Charles Loop, 1987). Its refine step splits every triangle into 4 by connecting the midpoints of its three edges; its smooth step then repositions points with triangle-specific masks:

New edge vertex: for an interior edge, take \tfrac{3}{8} of each of the edge's two endpoints plus \tfrac{1}{8} of each of the two opposite vertices of the adjacent triangles.
Updated original vertex: blend the old vertex with the average of its n neighbours using a valence-dependent weight \beta.

Loop and Catmull–Clark are two answers to the same question. Both give C^2 limit surfaces almost everywhere; both drop to C^1 at a handful of irregular points. The only real difference is the primitive — triangles vs quads — which is why a modeller chooses the scheme to match how they like to build topology.

A Catmull–Clark mesh is regular where every vertex has valence 4 (four quads meeting). Anywhere the topology forces a different valence — a triangle in a quad mesh, three or five quads meeting at a point — you get an extraordinary vertex. (For Loop's triangle meshes the regular valence is 6, and anything else is extraordinary.) There is usually no way to avoid them: you cannot tile a sphere or a character's face with valence-4 quads everywhere.

The subtle catch: the limit surface is beautifully C^2 (curvature-continuous) across regular regions, but only C^1 — tangent-continuous, yet curvature-discontinuous — right at an extraordinary vertex. It never creases or tears, so the silhouette looks fine, but the curvature jumps there, which a raking light can reveal as a faint pinch or shading wobble. That is why experienced modellers keep meshes quad-dominant and push extraordinary vertices into flat, low-curvature regions where the artefact can't catch the light — never onto a cheekbone or a specular highlight.

Creases: dialling sharpness back in

Pure subdivision rounds everything, which is wrong for a table edge, a sword, or a clenched jaw. The fix is a crease: tag an edge (or chain of edges) as sharp, and the smoothing step uses a different, curve rule there — averaging only along the crease, not across it — so the surface stays creased instead of rounding off. Tag a whole loop and you get a hard edge; tag a single vertex and you get a spike.

Even better are semi-sharp creases (Pixar's DeRose et al., 1998, first used at scale on Geri's Game): give the crease a real-valued sharpness s, and the sharp rule is applied for the first s levels of subdivision and the ordinary smooth rule afterward. Fractional values interpolate. This turns sharpness into a single continuous dial — from razor edge to soft bevel — that an animator can even keyframe, so a character's brow can harden as they frown.

Why animators love it

Subdivision surfaces win because they cleanly separate the two things a production needs. The coarse cage is what humans touch: a few hundred vertices are easy to rig, fast to skin, and light enough to animate at interactive speed — the rig deforms only the cage. The smooth limit surface is what the camera sees, generated on demand and subdivided as finely as the shot requires (more near the lens, less far away). One control mesh drives a resolution-independent surface: a single asset that is cheap to pose and gorgeous to render, with no re-modelling between the two. That is why Catmull–Clark, not raw polygons or trimmed NURBS, became the quiet standard of feature animation.

Half of it. Ed Catmull co-published the Catmull–Clark scheme in 1978, long before Pixar existed — but Pixar turned it into the backbone of production, adding the practical machinery (semi-sharp creases, efficient evaluation, robust handling of extraordinary points) that made it usable on real films. Their short Geri's Game (1997) was the first showcase built entirely on subdivision surfaces, and the underlying library eventually became the open-source OpenSubdiv, now used across the whole industry. So: co-invented in academia, weaponised for film at Pixar.