Every object on screen — players, the football, the field, the crowd — is built from
thousands of 3D Gaussians: soft, translucent ellipsoids defined by a position,
covariance (shape), color, and opacity. Each player alone is composed of 30+ overlapping
Gaussians (helmet, visor, shoulder pads, jersey, numbers, legs, cleats, shadows). The full
scene renders 4,000+ Gaussians per frame at 60 fps.
4D = 3D + Time
Standard 3D Gaussian Splatting reconstructs static scenes. This engine adds a
temporal dimension: each Gaussian carries keyframes that define how its position,
shape, color, and opacity evolve over time. Arm swing during a sprint, the spiral rotation of a
thrown football, crowd wave motion — all are encoded as continuous temporal
interpolations across Gaussian parameters rather than traditional skeletal animation.
Rendering Pipeline
Each frame follows a strict pipeline:
Scene Graph → View-Projection Matrix → Perspective Divide
↓
Depth Sort (back-to-front) → Alpha Composite
↓
Post-Processing: Fog → Vignette → Film Grain → Color Grade
Gaussians are projected from world space into screen space using a perspective projection
matrix, then depth-sorted back-to-front for correct transparency. Each
Gaussian is drawn as a radial gradient (Canvas2D's closest approximation to a
true splatted ellipsoid). A post-processing stack applies atmospheric depth fog,
cinematic vignette, photographic film grain, and
warm color grading.
Three-Loop Architecture
FAST LOOP (~60 Hz) Render + physics + particle dynamics
MEDIUM LOOP (~10 Hz) Coherence gate + play state machine
SLOW LOOP (~1 Hz) Game governance + AI play-calling
The fast loop handles rendering, camera interpolation, and particle physics every
frame. The medium loop runs a coherence gate that validates world-state
self-consistency — are all 22 players accounted for? Is the ball on the correct trajectory?
The slow loop governs high-level game logic: selecting formations, calling plays,
managing clock and score state.
Play Simulation Engine
Plays are not scripted animations. The engine defines offensive formations
(Shotgun, I-Form, Spread) and defensive coverages (Nickel, Cover 3), then
simulates route-running, QB dropbacks, handoffs, and tackle detection using
distance-based collision and probabilistic outcomes. Each play generates unique
player trajectories, yardage, and game-state transitions.
Broadcast Presentation
The HUD replicates an ESPN-style scorebug with real-time quarter, clock, down &
distance, possession, and play clock. Five camera modes (Broadcast, End Zone, SkyCam,
Follow Ball, Cinematic Dolly) use lerped position smoothing for fluid transitions.
Touchdown confetti, turf spray particles, slow-motion replay at 0.25x, and big-play
overlays complete the broadcast experience.
Why Gaussians?
Gaussian Splatting has emerged as a breakthrough in neural scene representation. Unlike mesh-based
rendering, Gaussians naturally handle transparency, soft edges, and volumetric
effects (fog, glow, motion blur) without expensive ray-marching. They compose additively,
making it trivial to layer complex appearances from simple primitives. This engine demonstrates
that even without GPU shaders, the Gaussian paradigm produces visually rich, physically
plausible scenes in real time on a standard Canvas2D surface.