Splat II · Deep dive

Visual Effects

Volumetric fog from graphics papers, a trail renderer driven by acceleration, and a 27-uniform black hole shader.

Three effects do most of the visual work in Splat II. All three are built from scratch.

Volumetric fog

Screen-space ray marching with Henyey-Greenstein scattering. Researched from graphics papers, rewritten four times to hit 60 fps.

How a pixel gets its colour

camera
   │
   ▼
┌──────┐
│ ray  │ (one per pixel)
└──┬───┘
   │
   ▼ 48 sample steps along the ray
   ●───●───●───●───●───●───●───●───●───●───●───●─...
   │   │   │   │   │   │   │   │   │   │   │
   ▼   ▼   ▼   ▼   ▼   ▼   ▼   ▼   ▼   ▼   ▼
  density at each sample
    = height term × ground proximity × 3D wind noise

  light at each sample
    = sum over lights of Henyey-Greenstein( angle_to_light )

  pixel = integrate density × light along ray, clamped by depth buffer

Sample count adapts from 16 (low) to 96 (ultra). 48 is the default.

Performance pass

Full resolution was too expensive. Fog is low-frequency so the trick is to render small and reconstruct:

half-resolution fog buffer
   │
   ▼
Halton-jittered ray origins (cuts visible banding)
   │
   ▼
temporal filter (blend with reprojected previous frame)
   │
   ▼
upscale to full screen
   │
   ▼
composite into the scene

Player lights scatter inside it

Each player emits a point light. The fog samples those lights at every step, so a player on the other side of the map shows up as a glow inside the fog even when their body is hidden behind geometry. Turned out to be one of the strongest readability cues in playtests.

Iteration history

Version	What it was	Why it failed
v1	Basic exponential fog	Too flat, no shape
v2	Layered fog at heights	Visible banding between layers
v3	Thin fog film overlay	No depth, looked like a filter
v4	Full ray-marched volumetric	Shipped version

Each version taught me what the next one had to fix.

Trail renderer

A glowing ribbon behind each player. Ring buffer of points, rebuilt into a triangle strip every frame on the CPU. Old points expire on a lifetime so the buffer stays bounded.

   [P0][P1][P2][P3][P4][P5][P6][P7][P8] ... [PN]
    ↑                                          ↑
    oldest point                       newest point (player pos)

each frame:
    push player position
    drop points older than lifetime
    rebuild triangle strip mesh from the live window

Acceleration drives the glow

Speed is roughly constant most of the time and makes a bad input. Acceleration spikes on boosts, grapple yanks, and direction changes, so driving the glow off acceleration makes those moments visibly flash.

glow_intensity = base + k * length(acceleration)

The shader builds a solid core stripe with soft edges plus a wider glow halo. Smooth turn handling prevents kinks where the path bends sharply.

Tuning history

First version covered the screen and obscured the city. I iterated the width curve, alpha falloff, and segment length until the trail read as a path instead of a banner.

Black hole shader

Replaced the original basic goal object. Fully procedural, no textures.

27 uniforms, grouped

Group	Controls
Spiral arms	count, angle, twist rate, density
Halftone dots	dot size, screen scale, threshold
Screen-space distortion	strength, radius, falloff curve
Colour cycling	palette, cycle speed, mix curve
Phase timing	activation, pulse rate, intro and outro

Around the shader

A physics sensor sphere detects when the player enters the gravity well and triggers the end-of-level event. An ambient particle system with a “towards point” affector swirls particles into the centre. The pulled particles double as a long-range visual cue: from across the map you can see where the goal is.

Workflow

I built the shader in Shadertoy first because iteration is fast there. Once it looked right, I ported it into the engine and tuned across a few commits to lock in the spiral animation and the distortion field.