Blendshape and Muscle Faces

A human face has around forty-three muscles, and a lifetime of watching other people use them has turned every one of us into a ruthless expert. We spot a fake smile across a room; we feel that a stiff jaw or a pair of dead eyes is wrong before we can say why. That expertise is wonderful for social life and brutal for the animator: the face is the one surface an audience will never let you fake. So how do you build a digital face that can act?

There are two great traditions, and this page sets them against each other. The first sculpts every expression by hand and blends between them — the blendshape (morph-target) face, the workhorse of film and games. The second builds a little anatomy — muscles pulling on soft tissue over a skull — and lets the expression emerge from physics: the muscle (anatomical) face. Each buys realism with a different currency, and the modern digital human quietly cheats by using both.

The blendshape face: sculpt, then blend

A blendshape rig starts from a single neutral mesh and a library of target meshes, each the same topology but sculpted into one expression — a raised brow, a squinted eye, a corner-puller for a smile. This is exactly the morph-target machinery you have already met, pointed at a face. Each target i is stored as a delta from neutral, \mathbf{d}_i = \mathbf{B}_i - \mathbf{N}, and the final face is neutral plus a weighted sum of deltas:

\mathbf{F} \;=\; \mathbf{N} \;+\; \sum_{i} w_i\,\mathbf{d}_i, \qquad w_i \in [0,1].

The targets are usually organised around the FACS Action Units — the anatomically-grounded catalogue of the smallest visible facial movements — so a rig might carry one shape per Action Unit plus a pile of combination shapes. Drive the weights from sliders, from motion capture, or from a machine-learning solver and the face performs.

The face is a linear combination of hand-sculpted expression targets: \mathbf{F} = \mathbf{N} + \sum_i w_i \mathbf{d}_i.
Strengths: fully art-directable (the artist sculpts the exact shapes they want), cheap to evaluate in real time, and every pose is reachable and predictable.
Weaknesses: blending is linear, so combinations can self-intersect or deflate, and fixing that needs many hand-sculpted corrective shapes — a combinatorial explosion.

The muscle face: simulate, and let expression emerge

The anatomical camp, pioneered by Keith Waters in 1987 and pushed to full soft-tissue simulation by researchers like Eftychios Sifakis and the film labs at Pixar and Disney, throws out the library of poses. Instead it builds a skull, a set of muscles anchored to it, and a layer of soft tissue (fat and skin) draped over the top. You do not animate shapes; you animate muscle activations a_j \in [0,1]. Each activation makes a muscle contract, which pulls on the tissue, which is solved as a deformable continuum so the skin bulges, slides, and folds.

The pay-off is that the physically correct behaviour comes for free: skin slides over bone rather than stretching like rubber, tissue conserves volume so a squeeze in one place bulges another, and wrinkles appear where compressed skin has nowhere to go. Nobody sculpted those wrinkles; they are what the simulation does. The cost is equally honest: a soft-tissue solve is far more expensive than a weighted sum, and "make this exact charming smirk" is genuinely hard when your only knobs are muscle contractions.

Expressions emerge from muscle activations a_j driving a physical simulation of soft tissue over a skull — not from stored shapes.
Strengths: physically plausible sliding, volume preservation and wrinkles; far fewer hand-sculpts; behaviour generalises to poses no one authored.
Weaknesses: computationally expensive, and hard to art-direct to a precise target — you steer muscles, not the final shape.

Feel the weights

Here is a toy face built the blendshape way: a neutral mesh plus a smile target (mouth corners up and out) and a jaw-open target (chin drops, mouth opens). Drag the two weights and watch the face morph as a live weighted sum. At full smile the corners lift; at full jaw the mouth yawns; turn both up and you get a laughing open smile — but look closely at how the additive shapes stack, because that combination is exactly where naïve blending gets into trouble.

Every vertex you see is \mathbf{N} + w_\text{smile}\,\mathbf{d}_\text{smile} + w_\text{jaw}\,\mathbf{d}_\text{jaw} — the same linear formula, evaluated per point, per frame.

Worked example: the smile-plus-jaw collision

Consider a single vertex on the upper lip. The smile target lifts it and curls it inward; the jaw-open target, which was sculpted with a closed mouth, leaves the upper lip roughly where it was. Take deltas (in millimetres) of \mathbf{d}_\text{smile} = (2,\,4,\,-3) and \mathbf{d}_\text{jaw} = (0,\,-1,\,1) from a neutral point at the origin. Pure additive blending at w_\text{smile} = w_\text{jaw} = 1 puts that vertex at

\mathbf{F} = (0,0,0) + (2,4,-3) + (0,-1,1) = (2,\,3,\,-2).

The z-coordinate is now -2\,\text{mm} — the lip has been driven behind the lower teeth, because the smile's inward curl was authored assuming a closed jaw. With the jaw dropped, that curl now pushes the lip through the mouth interior. This is a self-intersection, and it is not a bug in the maths — the linear sum did exactly what it promised. It is a bug in the assumption that two independently sculpted shapes can be added.

The blendshape fix is a corrective (combination) shape \mathbf{c} that fires only when both weights are high, typically gated by their product:

\mathbf{F} = \mathbf{N} + \sum_i w_i\,\mathbf{d}_i \;+\; (w_\text{smile}\,w_\text{jaw})\,\mathbf{c}.

The artist sculpts \mathbf{c} to push the lip back out to where a real laughing mouth sits — say \mathbf{c} = (0,0,3), cancelling the intrusion. Correct, but note what just happened: this one interaction needed a hand-sculpted patch. With n primary shapes there are \binom{n}{2} pairwise combinations, plus triples, plus… A production face rig can carry hundreds of correctives just to keep combinations from breaking.

The muscle model gets this for free. There is no "smile shape" to collide with a "jaw shape". You activate the zygomaticus major (the smile muscle) and drop the jaw joint; the soft-tissue solve moves the lip along the surface of the teeth because collision and volume preservation are in the simulation. The laughing mouth is simply what those two activations produce together — no corrective, because nothing was ever additively stacked in the first place.

Because the two paradigms fail in opposite places, the modern answer is to combine them. A typical high-end film or digital-human rig uses blendshapes for the primary shapes — where art direction matters most and artists want to sculpt exact poses — and layers physics on top: a thin soft-tissue or skin-sliding solve to add secondary motion (jiggle, sliding over bone) that blendshapes cannot express, plus wrinkle maps — normal-map textures blended in by expression that fake fine creasing without extra geometry. You get the art-directability of blendshapes and much of the physical plausibility of simulation, at a cost the pipeline can actually afford. Systems like Epic's MetaHuman lean on exactly this blendshape-plus-corrective-plus-map recipe.

The intuition "more realistic = better" is a trap here. As a face climbs toward photorealism it enters the uncanny valley: the closer it gets to human, the more a small error reads not as "a bit off" but as creepy — dead, corpse-like, wrong. A cartoon with no micro-expressions is fine, because you never expected them; a near-perfect digital human with slightly dead eyes, skin that doesn't slide right, or a smile that misses the eyes reads as unwell. The realism you added is precisely what raises the audience's expectations and makes the remaining flaw scream. This is why a deliberately stylised face often reads as more alive than a 99%-perfect one, and why landmark efforts like Digital Emily spent enormous care on the eyes, skin scattering and micro-motion — the exact places a nearly-real face betrays itself.

The blendshape-specific twist: pushing for realism means more targets and more combinations, and additive blending breaks combinations. So the drive toward photoreal simultaneously (a) makes every small facial error more disturbing, and (b) multiplies the number of corrective shapes you must sculpt to stop the face from breaking. Realism raises the bar from both sides at once.

Which paradigm, when?

	Blendshape (morph-target)	Muscle (anatomical)
You control	expression weights w_i	muscle activations a_j
Expression comes from	hand-sculpted targets	physical simulation
Art-directability	excellent — exact shapes	indirect — steer muscles
Combinations	need corrective shapes	emerge naturally
Wrinkles / sliding / volume	faked (maps, correctives)	physical, for free
Cost	cheap (a weighted sum)	expensive (a solve)
Typical home	games, real-time, film primaries	research, hero close-ups, secondary layers

In practice the honest answer is "both": blendshapes for the shapes an artist must nail, a physical layer for the motion no one wants to sculpt, and wrinkle maps to paper over the fine detail. The two traditions did not so much compete as merge.