You spin the camera toward the upper deck. Fifty thousand fans, all cheering. Your frame rate doesn't flinch. Take a breath and sit with that for a second, because what just happened is one of the more elegant cons in software engineering.
The engine is deciding, right now, which of those people actually exist.
The answer runs through a tiered priority system called a Level of Detail (LOD) hierarchy, applied not just to meshes but to the simulation itself. Distance is the first filter. What you're looking at is the second. Processing budget is the third. Everything else is theater.
The three tiers your GPU never tells you about
Picture a football crowd in a game like EA FC or Madden. The engine carves the visible world into concentric zones around the camera.
In the inner zone, roughly thirty meters out, you get full NPCs: individual skeletons, blend-shape facial animations, cloth physics on scarves, and a behavior state machine tracking whether a character is sitting, standing, or reacting to a goal. These are the people who'll catch your eye in a replay cutscene. The engine is running real pathfinding and reaction logic on them, roughly equivalent to what it spends on enemies in a combat game.
Pull back to the middle tier, sixty to a hundred meters, and the simulation quietly swaps. Skeletal animation compresses to a cheaper two-bone approximation. The state machine collapses to a handful of looping clips: idle, cheer, sit. Cloth physics disappears entirely.
The character still looks like a person. It is not being treated like one.
Beyond that, individual characters often stop existing at all. The engine replaces them with an impostor: a flat, camera-facing sprite, or a pre-baked texture atlas of crowd patches cycling through a handful of frames. No simulation running. It's wallpaper that breathes.
Red Dead Redemption 2's towns make this concrete. Residents close to Arthur Morgan run full AI schedules: they shop, go home, react to crime. The person across the street is running a stripped behavior tree. The figures visible through a saloon window in the background? A looping animation clip, no pathfinding, no world awareness whatsoever. Rockstar has discussed this in tech breakdowns, and the seams are invisible unless you know exactly where to look. That invisibility is the whole achievement.
Why distance alone isn't enough
Here's the counterintuitive part. Distance isn't the only trigger.
Engines also track screen-space coverage, basically how many pixels a character is actually occupying at a given moment. A character forty meters away in an open field might cover four hundred pixels and get mid-tier simulation. The same character forty meters away at the end of a narrow alley, framed tightly by walls, might cover two thousand pixels and get promoted to full simulation on the spot. The engine is asking: will the player notice if this is cheap? Pixel count is the proxy for that question, and it's a smarter proxy than pure distance ever could be.
On top of that, most modern engines run an importance budget, a soft cap on how many fully-simulated NPCs can exist simultaneously. In Cyberpunk 2077, this budget shifts dynamically. Stand in a quiet alley and nearby pedestrians get promoted. Walk into a crowded market and some of those same people get demoted mid-scene, their AI quietly suspended while a pooled animation keeps them moving. CD Projekt Red documented several of these systems in post-launch patches when players noticed pedestrians behaving oddly at the edge of vision. The fix wasn't adding simulation. It was tuning where the cutoff lived.
So here's the question worth sitting with: if two players run the same open-world game on different hardware, are they actually playing the same city? One GPU affords a larger full-simulation radius. The other hits its budget threshold at half the distance. Same street corner, same crowd density, entirely different amounts of actual cognition happening inside those virtual skulls. Neither player notices. That is, genuinely, the point.
What people consistently misread about this
The common assumption is binary: crowd members are either fully simulated or purely cosmetic. That framing is wrong, and it undersells how clever the engineering actually is.
The reality is a continuous spectrum shifting in real time across three variables at once: your distance, your camera angle, and the current frame budget. An NPC can be promoted and demoted multiple times in a single scene. It's less like a light switch and more like a theater company that quietly swaps understudies every time the director looks away.
The transition between tiers is also rarely instantaneous. Engines use hysteresis, a deliberate lag before demotion, so a character you glance at briefly doesn't flicker between states. You'd notice that flicker. The slight delay is a feature disguised as inefficiency, which is maybe the purest definition of good UX.
The crowd is real where you're looking. Everywhere else, it's a precisely calibrated approximation running just enough simulation to survive your curiosity. The engineering isn't trying to simulate ten thousand people. It's trying to simulate ten thousand people for you, and that distinction is everything.