Loading field map
Every image generated by a computer is the product of a series of decisions about what to compute, how accurately to compute it, and how quickly to deliver the result. Graphics rendering—the subfield devoted to turning 3D scene descriptions into 2D images—has always been pulled between three competing pressures: the demand for photorealistic accuracy, the need for interactive speed, and the desire to communicate information clearly rather than realistically. The ten frameworks that structure the history of rendering each represent a different answer to these tensions, and the relationships among them—replacement, coexistence, absorption, and convergence—reveal how the field has evolved.
The earliest practical rendering systems were built on two frameworks that together defined the dominant approach for nearly two decades. Rasterization, which emerged around 1970, solved the fundamental visibility problem: given a set of 3D polygons, which ones are visible at each pixel? By projecting polygons onto the image plane and determining which fragments survive depth tests, rasterization provided a fast, hardware-friendly way to produce images. It did not, however, say anything about how those visible surfaces should be colored.
That task fell to Local Illumination, a shading framework that became the standard companion to rasterization from 1971 onward. Local illumination models—most famously the Phong reflection model—compute the color of a surface point using only the surface's material properties, the viewer direction, and the direct light sources. They ignore all indirect light bounces: light that reflects off one surface and onto another is simply not modeled. This approximation was computationally cheap and, for many applications, visually acceptable. The combination of rasterization for visibility and local illumination for shading became the backbone of real-time graphics for the next two decades, powering everything from CAD systems to early video games.
By the early 1980s, researchers recognized that local illumination's neglect of indirect light produced images that looked flat and artificial. Two competing frameworks emerged to address this limitation, each with a fundamentally different strategy.
Ray Tracing, introduced in 1980, simulates light by following rays backward from the camera into the scene. When a ray hits a surface, the algorithm spawns secondary rays toward light sources (for direct illumination) and in reflected or refracted directions (for indirect effects). The result captures sharp reflections, transparent shadows, and specular interreflections with striking realism. Ray tracing is a point-sampling method: it evaluates light transport at discrete ray–surface intersections, which makes it elegant for mirror-like reflections but computationally expensive for diffuse interreflection, where light scatters in many directions.
Radiosity, developed in 1984, took the opposite approach. Instead of sampling rays, radiosity treats the scene as a finite-element system: surfaces are subdivided into patches, and a system of linear equations computes the equilibrium distribution of diffuse light energy across all patches. Radiosity excels at capturing the soft, color-bleeding illumination that ray tracing struggles with—the warm glow of sunlight spreading across a matte wall, for instance. But radiosity assumes all surfaces are perfectly diffuse (Lambertian), and it cannot handle specular reflections or refraction at all.
For a time, ray tracing and radiosity coexisted as rival global-illumination paradigms, each with vocal advocates. Ray tracing was general but slow for diffuse effects; radiosity was fast for diffuse scenes but fundamentally limited to matte materials. Neither framework could claim universal superiority, and practitioners had to choose based on the scene content.
The rivalry between ray tracing and radiosity was resolved not by one side winning, but by a third framework that absorbed the strengths of both. Path Tracing, introduced in 1986, is a Monte Carlo method that generalizes ray tracing by using random sampling to approximate the full light transport integral. Instead of spawning a fixed set of secondary rays, path tracing shoots many random rays from each surface point, building up an estimate of the total incoming light through statistical averaging. This approach handles both diffuse and specular interreflections within a single, unified algorithm.
Path tracing is grounded in the Rendering Equation (1986), which formalizes light transport as an integral equation describing the equilibrium radiance leaving any surface point in any direction. The equation itself is framework-neutral—it can be solved by radiosity, ray tracing, or path tracing—but path tracing became the dominant solver because it makes no assumptions about surface reflectance. Radiosity, by contrast, solves only the diffuse subset of the Rendering Equation. As computing power grew and variance-reduction techniques improved, path tracing gradually displaced radiosity for offline rendering. By the late 1990s, radiosity had largely been absorbed into the broader Monte Carlo toolkit, and its active development as a standalone framework declined.
While the global-illumination debate unfolded, two other frameworks challenged the assumption that rendering should be about surfaces and realism.
Volumetric Rendering, which emerged around 1988, broke decisively with the surface-centric tradition. Instead of modeling objects as opaque boundaries, volumetric rendering treats scenes as distributions of participating media—fog, smoke, fire, clouds, or medical scan data—where light is absorbed, scattered, and emitted at every point in a 3D grid. The canonical approach, ray marching, steps along each viewing ray and accumulates color and opacity from the volume data. Volumetric rendering did not replace surface rendering; rather, it carved out a complementary domain where surfaces are meaningless. Today, volumetric methods are routinely combined with surface-based frameworks: a game engine might rasterize solid objects while ray-marching volumetric fog, and a production renderer might path-trace both surfaces and volumes in the same image.
Non-Photorealistic Rendering (NPR), which coalesced as a framework around 1990, rejected the premise that rendering's goal is physical accuracy. NPR researchers asked: what if the goal is to communicate form, emotion, or information, as in technical illustration, cartoon animation, or architectural sketching? NPR techniques deliberately distort lighting, simplify geometry, and add stylized outlines to produce images that are more legible or expressive than a photorealistic counterpart. NPR did not compete with physically based frameworks on their own terms; instead, it pluralized the very definition of what rendering is for. It remains an active research area, often borrowing infrastructure from rasterization and ray tracing while applying its own aesthetic criteria.
Two later frameworks further diversified the rendering landscape by questioning what counts as input and what counts as physically correct.
Image-Based Rendering (IBR), which began around 1995, proposed that novel views of a scene could be generated directly from a collection of photographs, bypassing geometric modeling entirely. The plenoptic function—the radiance arriving at every point in space from every direction—could be sampled by cameras and then resampled to produce new viewpoints. IBR offered a tantalizing shortcut: if you have enough photographs, you can render realistic images without building a 3D model. In practice, IBR never replaced geometry-based rendering because the required sampling density is prohibitive for general scenes, and because it struggles with dynamic lighting. Instead, IBR found a niche in image-based lighting (using environment maps to illuminate synthetic objects) and in light-field cameras. It coexists with geometric frameworks as a specialized tool rather than a general replacement.
Physically Based Rendering (PBR), which crystallized around 2004, is not a rendering algorithm but a material-shading discipline. PBR replaced the ad-hoc, artist-tuned parameters of local illumination with physically grounded models of surface reflectance—most commonly the Cook-Torrance microfacet model for specular reflection and the Lambertian or Oren-Nayar model for diffuse reflection. PBR's key commitments are energy conservation (a surface cannot reflect more light than it receives) and the use of physically meaningful parameters (roughness, metalness, index of refraction). PBR was rapidly adopted by the film industry and later by real-time game engines, where it provided a consistent, predictable shading language that artists could use across different lighting environments. PBR did not replace path tracing or rasterization; rather, it became the shading layer that both frameworks implement. A rasterizer using PBR materials produces more realistic images than one using Phong shading, and a path tracer using PBR materials converges to physically accurate results.
Real-Time Rendering is not a single algorithm but a research program defined by a hard constraint: the image must be generated within a few milliseconds to sustain interactivity. From 1980 onward, real-time rendering has been the primary driver of hardware innovation, from dedicated rasterization pipelines to programmable shaders to ray-tracing accelerators. For most of its history, real-time rendering meant rasterization plus local illumination, because those were the only frameworks fast enough. Over the past decade, however, real-time rendering has absorbed elements of path tracing (through denoised Monte Carlo sampling) and PBR (through standardized material models). Modern game engines like Unreal Engine 5 use a hybrid approach: rasterization for primary visibility, screen-space ray tracing for reflections and shadows, and path-traced global illumination for diffuse bounces. Real-time rendering thus acts as an integrating framework, pulling techniques from slower frameworks into the interactive domain through approximation and hardware co-design.
Today, no single framework dominates rendering. Instead, the field operates with a clear division of labor. Path tracing is the gold standard for offline rendering in film and architectural visualization, where accuracy matters more than speed. Rasterization remains the workhorse of real-time graphics, but it is increasingly augmented by ray tracing for specific effects (shadows, reflections, ambient occlusion) and by PBR for material shading. Volumetric rendering handles phenomena that surfaces cannot represent, and it is integrated into both offline and real-time pipelines. NPR serves applications where style or clarity is paramount, from medical illustration to animated films. Image-Based Rendering persists in specialized roles like environment lighting and light-field capture.
The leading frameworks today agree on one fundamental principle: the Rendering Equation is the correct theoretical foundation for light transport, and Monte Carlo methods are the most general way to solve it. They disagree on how much approximation is acceptable. Real-time practitioners accept bias and noise in exchange for speed, while offline renderers minimize bias at the cost of computation. A second disagreement concerns the role of geometry: surface-based frameworks (rasterization, ray tracing) assume that scenes are composed of discrete objects, while volumetric and image-based frameworks treat scenes as continuous fields or sampled light distributions. These disagreements are not signs of fragmentation; they reflect a mature field that has learned to match frameworks to tasks.
The history of graphics rendering is not a story of one framework triumphing over all others. It is a story of frameworks that replaced predecessors in specific domains (path tracing absorbing radiosity), coexisted as complementary tools (volumetric and surface rendering), converged into hybrid pipelines (real-time rendering absorbing path tracing and PBR), and pluralized the very goals of the field (NPR). The student who understands these relationships understands rendering.