Keyframing vs Simulation

Every moving image a computer produces was made by one of two great methods, and the choice between them shapes the whole production. In the first, an animator sits down and authors the motion by hand — posing a character at a few important moments and letting the machine fill the gaps. In the second, the animator authors almost nothing about the motion: they set up a little world — masses, forces, initial velocities — and press play, and the machine solves the physics to discover what happens. The first is keyframing; the second is simulation. Nearly everything that moves in a modern film is a blend of the two.

The tension between them is the whole story of this page: keyframing gives you total artistic control but makes you responsible for every wiggle, while simulation gives you physical plausibility almost for free but fights you when you try to direct it. Great pipelines don't pick a side — they use each where it is strong. Understanding which is which, for any given piece of motion, is one of the first instincts a technical animator has to build.

Keyframing: the animator drives

The animator authors sparse poses ("keys") at chosen key times — a channel's value is pinned at a handful of frames.
The computer interpolates the in-betweens along a F-curve connecting the keys, so the motion is fully determined by the keys and their tangents.
Every detail of the result is authored: nothing appears that the animator did not put there. This is total control — and total responsibility.

Keyframing is the direct descendant of hand-drawn animation, where a lead artist drew the extreme poses and assistants drew the frames between them. The value on frame t is just an interpolation of the surrounding keys — for two keys at times t_0, t_1 with a normalised, eased parameter u = e\!\left(\tfrac{t-t_0}{t_1-t_0}\right):

x(t) = (1-u)\,x_0 + u\,x_1, \qquad u \in [0,1].

Because you choose the keys, the eased shape e(\cdot), and the tangents, keyframing is the tool of choice whenever the motion has to mean something: a character's performance, an acting beat, a precisely timed comic pause. The computer never surprises you — which is exactly why hand-animating thousands of independent things is a nightmare.

Simulation: the physics drives

The animator specifies initial conditions (positions, velocities) and forces / constraints (gravity, wind, springs, collisions), not the motion itself.
A solver integrates the equations of motion forward in time, discovering the trajectory step by step from the current state.
The result is physically plausible by construction — but only indirectly directable: you steer it through the setup, not by touching frames.

At its heart a simulation just marches Newton's second law forward. Given a state (\mathbf{x}, \mathbf{v}) and net force \mathbf{F} on a body of mass m, one explicit time step of size \Delta t reads:

\mathbf{v}_{n+1} = \mathbf{v}_n + \frac{\mathbf{F}}{m}\,\Delta t, \qquad \mathbf{x}_{n+1} = \mathbf{x}_n + \mathbf{v}_{n+1}\,\Delta t.

Run that update for every one of a cloth's ten thousand vertices, coupled by spring forces and collision constraints, and you get drape and folds no human could pose by hand. This is why simulation owns cloth, hair, fluids, smoke, fire, rigid-body destruction, and secondary motion — the jiggle and follow-through of things that react rather than act. The cost is directability: to make the flag wave this way, you cannot grab a frame; you must change the wind, the mass, the stiffness, and re-solve.

The trade-off, on one chart

Plot the methods on two axes — how directable the motion is versus how physically plausible it comes out — and the landscape is clear. Keyframing lives in the top-left: maximum control, plausibility only if the animator earns it. Simulation lives in the bottom-right: plausibility for free, control only through the setup. The prize spot is the top-right, and the whole craft of production is dragging methods toward it: art-directing a sim to gain control, or referencing physics while keyframing to gain plausibility.

Notice there is a third point on the map. Not every non-keyframed motion is a physics solve.

A third category: procedural animation

Between "hand-posed" and "physics-solved" sits procedural animation: motion generated by a rule or algorithm that is neither a keyed curve nor a physical simulation. A cyclic \sin wave driving a fish's tail, a noise function jittering a flame, a crowd system where each agent follows steering rules, a walk cycle synthesised from procedural footstep planning — all are authored as logic, not poses or forces. Procedural methods are cheap, tunable and repeatable, and they scale to many instances, which is why they power crowds, foliage in wind, and background life. They occupy the middle of our chart: more automatic than keyframing, more directable than a raw simulation.

The honest picture, then, is three paradigms — keyframed, simulated, procedural — and real shots mix all three.

Directing a simulation without hand-animating it

"Uncontrollable" is too strong. A large part of effects work is art-directed simulation: bending a physics solve toward what the director wants while keeping it looking solved rather than posed. The main levers:

Tuning the setup — mass, stiffness, damping, wind strength, drag; the honest, physical way to change the outcome.
Guides and targets — a keyframed "guide" curve or goal shape the sim is pulled toward (guide hairs steering a fur groom; a target mesh the cloth is nudged to match).
Control forces — extra, non-physical forces added purely for direction: a gentle field that shapes smoke into a logo, a force that keeps a splash inside frame.
Constraints and pinning — nailing selected points (pin the cape at the collar) so the solve runs free everywhere else.

Each of these trades a little plausibility for a little control — nudging the simulation up-and-left on our chart, toward that coveted top-right corner.

The hybrid workflow: keyed primary, simulated secondary

The dominant workflow in character work is a hybrid: keyframe the primary motion — the intentional performance, the body acting — then simulate the secondary motion that must react to it. The classic example is a superhero: an animator keyframes the character's run and leap for a precise, readable performance, and a cloth simulation drives the cape, which flows and snaps in response to that keyed motion. Hair, cloth, jiggle, dangling props, antenna-wobble — all secondary, all simulated, all driven by the keyed body they hang off.

This is the trade-off resolved by division of labour: control where control matters (the acting) and plausibility where hand-work would be hopeless (ten thousand cloth vertices). The keyed rig becomes a moving boundary condition for the sim, which is exactly why the two paradigms are complements, not rivals.

Worked example: keyframe or simulate?

For each shot element, ask two questions: does the motion carry authored intent? (→ keyframe) and does it involve many reactive degrees of freedom obeying physics? (→ simulate). Here is the reasoning for five classic cases.

Element	Choice	Why
A hero's walk	Keyframe	The performance is the point — weight, attitude, personality. Every step is authored intent; a sim can't act.
A flag in wind	Simulate	Thousands of coupled cloth vertices reacting to wind and gravity — physically driven, no authored meaning, impossible to hand-pose convincingly.
An explosion / debris	Simulate	Fluid smoke plus rigid-body destruction: enormous numbers of interacting pieces obeying physics; hand-animation is out of the question.
An eye blink	Keyframe	A tiny, precisely-timed acting beat — its meaning lives in exact timing and spacing. Simulating it would be absurd overkill and lose the intent.
The hero's cape during the walk	Simulate (secondary)	Reactive cloth driven by the keyframed body — the canonical hybrid: keyed primary, simulated secondary.

The pattern falls out cleanly: intent and performance → keyframe; many reactive DOFs obeying physics → simulate; and reactive things hanging off a performance → keyframe the performance and simulate the reaction.

Largely by procedural and agent-based systems rather than physics. The battle scenes that made crowd tech famous used software where each soldier was an autonomous agent running a little brain — sense neighbours, pick a move from a library of keyframed clips, blend between them — so tens of thousands of characters could act plausibly without a human posing any of them, and without a full physics solve either. It is a perfect illustration that the world is not just keyframe-vs-sim: procedural logic is a genuine third paradigm, and big shots lean on it hard. The individual clips each agent plays back are keyframed; the choice and blending of those clips is procedural; and the stumble when two soldiers collide might be a rag-doll simulation. Three paradigms, one shot.

The two great beginner mistakes are mirror images of each other. The first is keyframing what should be simulated: nobody can hand-animate ten thousand cloth vertices or a plume of smoke into looking natural — the degrees of freedom are too many and the physics too subtle, so the result reads as stiff, dead, and wrong no matter how many hours you pour in. Reach for a solver. The second is simulating what should be keyframed: trying to get a physics sim to deliver a nuanced acting beat — a knowing glance, a comic double-take, a precisely weighted pause — is fighting the tool, because a sim has no intent. It will give you plausible but meaningless motion. The rule of thumb: if the motion must carry authored meaning, keyframe it; if it is many things reacting to physics, simulate it. Most shots need both, on different elements — and knowing which knob turns which is the skill.