Computer graphics is the branch of computer science concerned with representing visual information and producing images through computation. It includes two-dimensional drawing, three-dimensional modeling, image synthesis, animation, and interactive visual systems. Its outputs range from diagrams and screen interfaces to engineering models, scientific visualizations, and photorealistic film imagery. The field combines geometry, numerical methods, optics, perception, software design, and specialized computing hardware.
A central distinction is between describing a scene and generating an image of it. A scene may contain geometric objects, materials, lights, cameras, and motion; rendering converts that description into an image. Some systems instead work directly with pixels, curves, or other two-dimensional elements. Computer graphics is closely related to image processing, which transforms existing images, and computer vision, which infers information about a scene from observations. Their boundaries increasingly overlap, particularly in reconstruction, computational photography, and learned image synthesis.
The field serves both communication and simulation. A transit map emphasizes selected relationships rather than photographic appearance, while a physically based renderer approximates how light interacts with surfaces. Interactive applications must respond within strict time limits; offline production can spend minutes or hours computing a single frame. These different objectives explain the diversity of representations, algorithms, and engineering methods used in graphics.
Historical development
Computer graphics emerged from several partly independent activities: scientific plotting, military display systems, industrial design, and experiments in computer art. Early computers commonly produced numerical tables or printed characters, but oscilloscopes and cathode-ray-tube displays made it possible to draw points and lines electronically. In the 1950s, systems associated with the Whirlwind computer and the SAGE air-defense project demonstrated interactive graphical displays and light-pen input. Plotters offered another route to computer-generated images by moving pens over paper under numerical control.
In 1963, Ivan Sutherland presented Sketchpad, developed at the Massachusetts Institute of Technology. Its importance lay not merely in displaying drawings but in allowing users to construct and modify them interactively. Geometric constraints could keep lines perpendicular or preserve relationships between objects, while reusable drawing components anticipated later modeling and interface techniques. During the same period, industrial work on numerical control and automobile body design encouraged mathematical representations of smooth curves and surfaces. Pierre Bézier at Renault and Paul de Casteljau at Citroën developed influential methods for controlling polynomial curves.
Research during the 1960s and 1970s established many foundations of three-dimensional image synthesis. Work at the University of Utah and elsewhere addressed hidden surfaces, perspective, smooth shading, and curved objects. Henri Gouraud introduced an influential interpolation-based shading method in 1971, and Bui Tuong Phong developed a shading model and normal-interpolation technique in the early 1970s. Edwin Catmull’s 1974 doctoral work helped establish texture mapping and subdivision-based rendering. These developments made images of curved, illuminated objects practical without explicitly modeling every visible detail.
During the 1980s, workstations, personal computers, and dedicated graphics hardware expanded access to graphical interfaces and three-dimensional visualization. Turner Whitted’s 1980 paper demonstrated a recursive ray-tracing framework for reflections, refractions, and shadows. James Kajiya introduced the rendering equation in 1986, providing a general formulation of light transport. In 1995, Toy Story became the first fully computer-animated feature film, illustrating the maturity of production pipelines rather than the invention of any single underlying technique.
Commodity graphics accelerators transformed interactive graphics during the 1990s and 2000s. Fixed-function hardware evolved into programmable processors, while graphics programming interfaces provided increasingly explicit control over their operation. More recently, hardware-assisted ray tracing, large-scale parallel computation, and learned representations have brought previously separate approaches into closer contact. Nevertheless, line drawing, rasterization, geometric modeling, and physically based simulation remain active components of contemporary systems rather than obsolete stages in a linear progression.
Mathematical foundations and coordinate systems
Graphics relies heavily on linear algebra, especially vectors, matrices, and transformations. A point specifies a location, while a vector commonly expresses a displacement, direction, or velocity. Dot products measure directional alignment and support lighting calculations; cross products in three dimensions construct perpendicular directions and oriented surface normals. Distinguishing these mathematical roles matters: translating an object changes its points but should not change the direction of its normals or velocities.
A scene is normally expressed through several coordinate systems. Object coordinates describe geometry relative to an object’s own origin. World coordinates place objects in a shared environment, and view coordinates describe positions relative to a virtual camera. Projection then maps the visible scene toward an image plane. Homogeneous coordinates make translations and projective transformations compatible with matrix multiplication. A three-dimensional point is represented by four coordinates, with Cartesian coordinates recovered by division by the final component when it is nonzero.
For a typical vertex transformation, the successive operations can be written as
where , , and are model, view, and projection matrices. Matrix multiplication is generally noncommutative, so changing the order changes the result. Rotating an object around its own center, for example, differs from rotating it around a distant world origin. Hierarchical transformation systems exploit this structure: a hand can inherit the motion of an arm, which in turn inherits the motion of a body.
Perspective projection makes objects appear smaller as their distance increases. In a simplified camera model, projected coordinates are proportional to and , subject to the chosen axis conventions. Orthographic projection omits this distance-dependent scaling and is useful for engineering drawings, maps, and some interface views. Neither projection is intrinsically more accurate; each encodes a different relationship between spatial geometry and the intended image.
Calculus and numerical analysis become essential when surfaces, motion, and light vary continuously. Derivatives describe tangents, surface orientation, and rates of change, while integration expresses accumulated illumination or volumetric effects. Rotations are often represented using quaternions, which support compact orientation storage and useful interpolation methods. Numerical precision remains a practical constraint: very large coordinates can make nearby points difficult to distinguish, nearly singular transformations can amplify errors, and transforming normals under nonuniform scaling requires the inverse transpose of the relevant linear transformation rather than the transformation used for points.
Geometric representations and modeling
A graphics system must choose how to represent the shape of an object. The most common general-purpose representation is a polygon mesh, consisting of vertices connected into edges and faces. Triangles are especially useful because three noncollinear points define a plane and because dedicated hardware can process them efficiently. A mesh also stores or references attributes such as texture coordinates, surface normals, material assignments, and deformation weights. The visible quality of a mesh depends on its structure and attributes as well as its polygon count.
Connectivity carries information that positions alone do not provide. Neighboring triangles may share vertices and edges, allowing algorithms to traverse a surface, estimate curvature, or smooth selected regions. A manifold surface has a locally well-behaved neighborhood resembling a disk, or a half-disk at a boundary. Nonmanifold edges, self-intersections, and inconsistent face orientation can be acceptable for display but problematic for simulation, manufacturing, or volume calculations. Consequently, a model suitable for a film shot is not necessarily suitable for engineering analysis.
Smooth shapes can be represented parametrically. Bézier curves use control points to define polynomial curves, while nonuniform rational B-splines support more general piecewise rational curves and surfaces, including exact representations of important conic shapes. Such representations are central to industrial design because designers can control smoothness and continuity without manipulating every final polygon. Rendering often converts these surfaces into triangles through tessellation, with finer subdivisions where curvature or projected size requires them.
Subdivision surfaces bridge polygonal modeling and smooth-surface design. Repeated refinement of a coarse control mesh approaches a limiting surface, allowing artists to manipulate a manageable structure while obtaining smooth output. Catmull–Clark subdivision, introduced in 1978, is widely associated with this approach. Unlike simply smoothing vertex normals, subdivision changes the underlying geometric surface. Sharp features can be preserved through crease rules or carefully arranged control geometry.
Other representations suit different tasks. Constructive solid geometry builds objects through operations such as union, intersection, and difference on simpler solids. Implicit surfaces define shapes as level sets of functions, while signed distance functions encode distance to a boundary with a sign distinguishing interior from exterior. Voxels partition space into volumetric cells and are useful for scanned data, material fields, and some simulation methods. Point clouds store sampled positions without necessarily specifying connectivity; photogrammetry and depth sensing frequently produce such data. These alternatives trade compactness, editability, topological flexibility, and rendering efficiency rather than forming a single hierarchy of quality.
Two-dimensional graphics, typography, and compositing
Two-dimensional graphics describes images using either sampled picture elements or geometric drawing instructions. A raster image stores values on a pixel grid, whereas vector graphics describes objects such as paths, curves, fills, and strokes. A raster photograph has a fixed sampling resolution; a vector logo can be rendered at many resolutions without first enlarging a fixed pixel array. Vector content nevertheless becomes sampled when displayed on a conventional screen, so its apparent quality still depends on rasterization and filtering.
Paths provide a compact language for drawing. Line segments and polynomial curves form contours, which can be filled according to rules such as nonzero winding or even–odd inclusion. Strokes introduce additional choices: width, end caps, joins, and dash patterns. These choices are geometrically significant. A wide stroke around a tight corner may create spikes unless its join behavior is limited, and a mathematically precise hairline may be unstable when it falls between display samples.
Typography is a specialized graphics problem involving both geometry and language. Fonts commonly describe glyph outlines using curves, but text rendering also requires shaping: selecting glyphs, positioning combining marks, applying ligatures, and handling context-sensitive forms. Unicode identifies characters and related text properties but does not by itself specify the final glyph arrangement. At small sizes, hinting and careful sampling can preserve legibility when a character’s strokes occupy only a few pixels. Subpixel techniques can exploit a display’s physical color-element arrangement, although their assumptions may fail under rotation, scaling, or image reuse.
Layered images are combined through alpha compositing. Alpha usually represents opacity or fractional coverage under a specified model, rather than an additional color channel with arbitrary meaning. With premultiplied colors, placing a foreground over a background gives
and
The color values in these equations are already multiplied by their corresponding alpha values. Confusing premultiplied and unassociated representations can produce dark or bright fringes around translucent edges.
Compositing extends beyond simple layering. Masks restrict operations spatially; clipping paths confine drawing to geometric regions; blending modes combine colors for illustrative or photographic effects. Image formats impose further constraints: PNG supports lossless raster storage and transparency, while JPEG is commonly used for photographic images with lossy compression. Two-dimensional graphics remains fundamental to interfaces, publishing, cartography, and technical illustration, even when those outputs incorporate imagery rendered from three-dimensional scenes.
Rasterization, visibility, and the rendering pipeline
Rasterization converts projected geometric primitives into contributions to a discrete image. A conventional three-dimensional pipeline begins with geometry and attributes, transforms vertices into camera-related coordinates, clips primitives against the viewing volume, and determines which image samples are covered. For each covered region, the system generates fragments carrying interpolated information. A fragment is a candidate contribution to an image location, not necessarily a final visible pixel.
Clipping is necessary before perspective division and rasterization because primitives may cross camera planes or extend outside the visible region. Back-face culling can discard surfaces oriented away from the viewer when the object and material permit it. Other forms of culling operate at larger scales: objects outside the viewing volume need not be submitted, and objects hidden behind other geometry may be omitted through occlusion tests. These operations reduce work but must be conservative if they are to avoid removing visible content.
The depth buffer, or z-buffer, solves visibility locally by storing a depth value for each sample. When a fragment arrives, its depth is compared with the stored value; an accepted fragment may update both depth and color. This allows opaque triangles to be processed without globally sorting them from front to back. However, finite depth precision can cause nearly coincident surfaces to alternate in visibility, producing z-fighting. Perspective projection distributes conventional depth precision unevenly, so near-plane placement and depth representation have substantial practical effects.
Surface attributes are interpolated across triangles. Direct linear interpolation in screen space is not generally correct for quantities defined over a perspective-projected surface. Perspective-correct interpolation accounts for homogeneous depth, preventing texture patterns from appearing to bend or swim incorrectly across polygons. Normal vectors also require care: interpolated normals are usually renormalized before lighting calculations. The distinction between geometric normals and shading normals allows a coarse mesh to appear smoothly curved without changing its silhouette.
Transparency complicates the otherwise local depth-buffer solution. Standard alpha blending is order-dependent, so transparent surfaces often require sorting or specialized order-independent approximations. Deferred shading stores surface properties in intermediate buffers and computes lighting afterward, whereas forward shading evaluates lighting while processing visible geometry. Clustered and tiled approaches organize lights by screen or spatial regions to reduce unnecessary work. These are architectural choices rather than mutually exclusive rendering philosophies: a single application may use different methods for opaque surfaces, transparent objects, particles, and reflections.
Light transport and physically based rendering
Image synthesis can be framed as computing how light reaches a camera after interacting with a scene. In geometric optics, light is approximated by rays traveling along straight segments between interactions. This approximation is effective when relevant structures are much larger than optical wavelengths, but it does not automatically capture diffraction or interference. A renderer also chooses whether to represent light spectrally or through a limited set of color channels, and whether time-dependent effects can be ignored.
A standard formulation is the rendering equation:
Here, outgoing radiance equals emitted radiance plus reflected incoming radiance integrated over the surface’s incident hemisphere. The function describes directional reflection, and the cosine factor accounts for projected area. Because incoming illumination depends on other surfaces, the equation couples the entire scene.
Ray tracing finds intersections between rays and geometric objects. Camera rays determine which surfaces are visible; shadow rays test whether a light is obstructed; further rays can follow reflection or transmission. Testing every ray against every primitive is usually too expensive. Spatial acceleration structures, especially a bounding volume hierarchy, group geometry within nested bounds so that large sets of primitives can be rejected together. Their construction quality and update cost matter greatly for moving scenes.
Path tracing estimates light transport by sampling sequences of scattering events. Rather than evaluating every possible direction at every surface, it chooses a limited number of directions and averages their weighted contributions. Monte Carlo methods make this feasible for high-dimensional integrals, but finite sampling produces noise. Under standard independent-sampling assumptions, error decreases proportionally to the inverse square root of the number of samples; obtaining roughly half the error therefore requires about four times as many samples. Importance sampling directs computation toward contributions likely to matter, while multiple importance sampling combines complementary sampling strategies.
Different transport algorithms handle difficult paths with varying efficiency. Small bright lights, glossy reflections, caustics, and illumination passing through many interfaces can be challenging for basic path tracing. Participating media add scattering and absorption within space rather than only at surfaces, producing fog, smoke, and underwater effects. Physically based rendering does not imply exact simulation: practical systems still simplify geometry, material structure, wavelength dependence, and light paths. Its defining feature is a coherent connection between the image calculation and a physical model, making results more predictable as lighting or viewing conditions change.
Materials, shading, and surface detail
A material model specifies how a surface interacts with light, while shading is the computation that applies material and illumination models at visible locations. These concepts are related but not identical. A material may define reflectance, roughness, transparency, and internal scattering; shading evaluates their consequences under particular lights and view directions. Older empirical models often combined diffuse and specular terms chosen for convenient control, whereas physically based models place stronger constraints on energy conservation and directional behavior.
The bidirectional reflectance distribution function, or BRDF, relates incident illumination to outgoing reflected radiance at a surface. Common microfacet models treat a rough surface as a statistical collection of tiny reflecting facets. A typical specular term combines a facet-normal distribution, Fresnel reflectance, and a factor accounting for masking and shadowing between facets. Roughness controls the spread of facet orientations rather than simply applying a blur to an otherwise finished image. Conductors and dielectrics require different reflectance behavior; metallic surfaces generally have wavelength-dependent reflection and do not behave like diffuse paint with an added white highlight.
Texture mapping supplies spatially varying attributes from images or procedural functions. A mesh commonly stores two-dimensional texture coordinates that associate surface points with locations in an image. The same mechanism can supply base color, roughness, opacity, emission, or other parameters. Seams arise where a curved surface is cut into a flat parameterization, and uneven parameterization can concentrate texture resolution in some areas while stretching it in others. Procedural textures avoid some storage constraints but still require filtering and often depend on carefully chosen coordinate spaces.
Normal mapping changes the orientation used for shading without changing the underlying surface boundary. Bump mapping derives a similar perturbation from a height field. Displacement mapping instead moves the geometric surface and can therefore alter silhouettes, shadows, and intersections if sufficient geometry is available. These techniques are not interchangeable. A brick wall represented only by a normal map can display convincing local highlights but remains geometrically flat when viewed at a grazing angle or when another object contacts it.
Many materials require more than surface reflection. Skin, wax, and marble exhibit subsurface scattering, in which light enters at one location and exits at another. Glass requires transmission and refraction; thin films can produce wavelength-dependent interference colors; hair and cloth have directional structures that strongly affect appearance. Material authoring also separates physical parameters from artistic controls. A production may deliberately alter roughness, shadowing, or reflectance to preserve visual readability, while engineering visualization may prioritize consistency with measured optical properties. The appropriate model depends on what distinctions the image must communicate.
Sampling, antialiasing, and color
An image is a sampled representation of a spatially and temporally varying signal. A pixel should ideally represent a filtered contribution over a finite region, rather than the value of an infinitesimal ray at its center. Insufficient sampling produces aliasing: high-frequency detail appears as false lower-frequency patterns, jagged edges, flicker, or unstable motion. Thin geometry, repeated textures, sharp highlights, and rapidly moving objects are common sources. Increasing resolution helps, but filtering and sampling strategy determine how effectively the available samples represent the image.
Antialiasing methods address different parts of the problem. Supersampling evaluates more image samples and combines them through a reconstruction filter. Multisample antialiasing improves geometric edge coverage without necessarily repeating every shading calculation. Texture filtering integrates over the surface region projected onto an image sample; mipmaps provide prefiltered versions of a texture at progressively lower resolutions. Anisotropic filtering improves cases in which the projected texture footprint is elongated, as with a road receding toward the horizon. Simply selecting a single sharper texture level can increase detail in a still image while causing severe motion flicker.
Temporal sampling introduces further trade-offs. Motion blur approximates the integration of light over a shutter interval, while temporal antialiasing combines information across frames. Reusing previous frames can reduce noise and stabilize edges, but only if the system correctly accounts for motion and changing visibility. Newly revealed surfaces lack valid history, and moving transparent effects may not have a single reliable motion vector. Excessive reliance on past information can produce ghosting or smear fine details. Similar issues arise in temporal reconstruction and denoising for ray-traced images.
Color values must be interpreted within a defined color space and transfer function. An RGB triplet does not specify a unique physical color without primaries, a white point, and encoding conventions. Many familiar image encodings are nonlinear, whereas physically meaningful addition and blending generally require linear-light values. Applying illumination calculations directly to nonlinear display-encoded values can darken blends and distort energy relationships. Color management transforms values between working, storage, and display spaces so that images remain consistent across different devices and production stages.
High-dynamic-range imaging represents brightness over a range wider than many conventional displays can reproduce. Tone mapping converts scene-referred values into a displayable range while managing contrast and highlight detail. It is not merely a numerical rescaling: local adaptation, viewing conditions, and artistic intent influence the desired result. Spectral rendering offers another distinction, representing wavelength-dependent behavior more explicitly than three-channel rendering. This can improve effects such as dispersion and complex color interactions, although the final image still has to be mapped to the capabilities of a display.
Animation, deformation, and physical simulation
Animation describes how a scene changes over time. The changing quantities may include object transforms, articulated poses, surface geometry, material parameters, lights, or camera settings. Keyframe animation stores selected values at selected times and interpolates between them. Linear interpolation is simple but can produce abrupt velocity changes; spline-based interpolation offers smoother motion and more precise control over timing. The same geometric path can communicate different actions depending on acceleration, pauses, and the spacing of keyframes.
Character animation commonly uses an articulated skeleton. A rig provides controls that connect an animator’s actions to joints, deformers, facial shapes, and other internal parameters. Skeletal animation associates mesh vertices with one or more bones, allowing joint motion to deform the visible surface. Linear blend skinning is efficient but can produce volume loss and twisting artifacts. Alternative deformation schemes and corrective shapes address these limitations. Inverse kinematics solves for joint configurations that place an end effector, such as a hand or foot, near a desired location; because many poses may satisfy the same target, additional constraints or preferences are required.
Motion capture records aspects of a performer’s movement using markers, cameras, inertial sensors, or combinations of these. Captured motion is not automatically suitable for a different digital character. Retargeting must account for different limb lengths, joint limits, and proportions, while cleanup removes noise and resolves missed observations. Facial animation may combine captured expressions with artist-authored shapes. Procedural animation supplies motion through rules, such as a walking system that adjusts foot placement to terrain or a crowd model that controls many individuals without individually keyframing each one.
Physical simulation generates motion by approximating dynamical laws. A basic particle model follows
where forces can include gravity, elasticity, drag, and contact responses. Cloth, deformable solids, fluids, and rigid bodies require different discretizations and constraints. Explicit time integration is often straightforward but may become unstable unless the time step is sufficiently small. Implicit methods can tolerate stiff forces better, although they require solving more complex systems. Collision detection and contact handling are major costs because objects must not unrealistically pass through one another.
Graphics simulation often balances physical fidelity against controllability and production needs. A fluid solver may be adjusted to preserve a desired silhouette, and a character’s clothing may require special handling to avoid visually distracting intersections. Determinism, reproducibility, and caching matter when scenes are revised or rendered across many machines. The animated state and the rendering process are also distinct: a simulation can be computed once and rendered repeatedly with different lighting, while motion blur may require evaluating the geometry at multiple times within each frame’s exposure interval.
Graphics hardware and programming interfaces
Graphics hardware is organized to process large amounts of similar work in parallel. A graphics processing unit typically executes many lightweight threads over vertices, fragments, rays, or general data elements. This differs from optimizing a single sequence of complex operations on a central processing unit. The distinction is not absolute, but graphics workloads benefit strongly from regular computation, high memory bandwidth, and the ability to keep many independent operations in flight.
A shader is a program used within a graphics computation, commonly to transform geometry, evaluate surface appearance, or process image data. Early accelerators exposed relatively fixed operations; programmable stages allowed developers to replace many of these with custom calculations. Modern systems can include vertex, fragment, compute, tessellation, geometry, mesh, and ray-tracing-related stages, depending on the hardware and interface. Despite the name, shaders are not limited to calculating shade: they may perform culling, particle updates, deformation, or data transformation.
Memory behavior is often as important as arithmetic throughput. Nearby threads that access nearby data can use memory transactions efficiently, while scattered accesses waste bandwidth. Branches that send neighboring execution lanes down different paths can reduce utilization. Textures, geometry buffers, acceleration structures, and intermediate images compete for storage and cache capacity. A scene with modest polygon counts can still be expensive if it uses many large textures, excessive transparency, or poorly organized material changes.
An application programming interface mediates between application software and the graphics system. OpenGL, Vulkan, and Direct3D represent different generations and approaches to exposing graphics capabilities. More explicit interfaces give applications greater responsibility for memory management, synchronization, and scheduling. This can reduce overhead and enable more predictable performance, but it also increases implementation complexity. Browser-based graphics adds another layer: WebGL and WebGPU expose capabilities within the constraints of a browser’s portability and security model.
Real-time performance depends on total frame latency, not just the speed of one rendering stage. At 60 frames per second, a frame interval is about 16.7 milliseconds; at 120 frames per second, it is about 8.3 milliseconds. Simulation, command preparation, rendering, synchronization, and presentation all draw on this budget. Profiling therefore examines bottlenecks across the entire pipeline. General-purpose GPU programming, including CUDA, shares hardware foundations with graphics but uses different execution abstractions. Specialized ray-intersection hardware accelerates selected operations; it does not remove the costs of material evaluation, memory access, scene construction, or light-transport sampling.
Scientific visualization and technical applications
Graphics becomes scientific visualization when images are designed to reveal the structure of measured or simulated data. The underlying information may be a scalar field such as temperature, a vector field such as fluid velocity, or a collection of geometric and statistical relationships. Unlike illustration whose main goal is appearance, visualization must preserve the meaning of the data transformations it applies. Choices of interpolation, scale, viewpoint, and color can expose useful patterns or conceal them.
Surface extraction converts volumetric measurements into geometric boundaries. Marching cubes, published by William Lorensen and Harvey Cline in 1987, constructs triangle approximations to an isosurface by examining values at the corners of grid cells. This is useful when a meaningful threshold defines a structure, but the chosen threshold can substantially affect the resulting shape. Direct volume rendering instead accumulates contributions through a three-dimensional field without first extracting a single boundary. A transfer function maps data values to color and opacity, making selected ranges visible while suppressing others.
Medical imaging provides an important application. Data from magnetic resonance imaging and other imaging modalities can be displayed as slices, surfaces, or volumes. These views support inspection of spatial relationships, but a rendered image depends on acquisition resolution, segmentation, interpolation, and display settings. A smooth-looking surface does not imply that the original measurement resolved all visible detail. For technical and clinical systems, preserving coordinate conventions, scale, and provenance is as important as producing a visually clear image.
Engineering uses graphics throughout design and analysis. Computer-aided design systems represent manufactured parts and assemblies, while simulation tools display stress, deformation, flow, and heat transfer. Geographic applications combine terrain, imagery, and spatial attributes through a geographic information system. Large spatial scales introduce challenges such as coordinate precision, map projection distortion, and the need to stream different resolutions of terrain. Scientific applications may similarly require out-of-core processing when a dataset exceeds available memory.
Data visualization also includes abstract quantities with no natural three-dimensional form. Graphs, plots, and diagrams use position, length, area, color, and connection to encode information. Adding a third dimension can help with genuinely spatial data but can also introduce occlusion and perspective distortion into otherwise simple comparisons. Uncertainty requires explicit attention: ensembles, confidence intervals, probabilistic fields, and missing data should not be silently converted into a single apparently exact surface. Reproducibility may require recording not only the source data but also camera parameters, filters, transfer functions, and interactive selections that produced a particular view.
Interaction, virtual environments, and augmented reality
Interactive graphics closes a feedback loop between user input, scene state, and visual output. A user moves a pointer, changes a parameter, or turns their head; the application updates its model and displays a response. The quality of this loop depends on latency, stability, and the relationship between the user’s action and its visible effect. A technically detailed image may still be difficult to use if selecting an object is unreliable or if the display responds too slowly.
Picking identifies which object or scene element corresponds to an input location. A system may cast a ray through a screen coordinate, query an acceleration structure, or render object identifiers into a hidden buffer. Manipulation then converts input into meaningful constraints: dragging can move an object within a plane, rotate it around an axis, or adjust a point on a surface. These operations require more than rendering because a two-dimensional device must control a three-dimensional state without leaving the intended interpretation ambiguous.
Virtual reality presents an environment that responds to tracked user motion, often through a head-mounted display. Separate images for the two eyes provide binocular disparity, while head tracking updates the viewpoint. Motion-to-photon latency includes sensing, pose estimation, simulation, rendering, and display response. Late-stage reprojection can update an already rendered image using a more recent pose, reducing some apparent delay without fully recomputing the scene. It cannot perfectly recover newly visible surfaces that were absent from the original image.
Augmented reality combines computer-generated content with a view of the physical environment. Correct placement requires tracking the device and estimating aspects of the surrounding scene. Convincing integration also depends on occlusion, scale, lighting, and focus cues. A virtual object should sometimes pass behind a real one, which requires knowledge of real-world depth. Optical see-through and camera-mediated systems face different constraints: one combines real light with display light directly, while the other processes a camera image before showing the combined result.
Immersive rendering introduces specialized efficiency techniques. Foveated rendering allocates more detail near the user’s gaze and less in peripheral regions, taking advantage of nonuniform visual acuity. Its effectiveness depends on gaze tracking, transition quality, and latency. Spatial audio, haptics, and physical controllers extend the experience beyond images, while accessibility considerations affect text size, contrast, control schemes, and motion presentation. Collaborative environments add network delay and consistency problems: different participants must perceive sufficiently compatible scene states even when updates arrive at different times.
Reconstruction, neural rendering, and generative methods
Traditional graphics often assumes that a scene description already exists. Reconstruction instead estimates geometry, appearance, motion, or illumination from observations. The task may use photographs, depth maps, video, or combinations of sensors. Multiple scene explanations can produce similar images, so reconstruction is generally an ill-posed problem. A dark region, for example, might result from dark material, weak illumination, shadow, or a change in surface orientation. Additional viewpoints, calibration, physical constraints, and prior assumptions help distinguish these possibilities.
Differentiable rendering connects image formation with mathematical optimization. If a renderer provides useful derivatives of image values with respect to scene parameters, an optimizer can adjust those parameters to reduce an image-based error. This approach can estimate shape, camera pose, materials, or lighting. Visibility discontinuities pose a major difficulty because a small geometric change can suddenly reveal or hide a surface. Practical methods therefore use specialized derivative estimators, approximations, or smoother alternative formulations.
Neural radiance fields, introduced in their widely influential form in 2020, represent a scene through a learned function that predicts density and view-dependent emitted color from spatial location and viewing direction. Images are synthesized by sampling this representation along camera rays and integrating their contributions. The method demonstrated strong novel-view synthesis from collections of calibrated photographs, but its representation is not automatically an explicit, editable surface. Training coverage, camera estimation, dynamic content, and difficult reflectance all affect reconstruction quality.
3D Gaussian splatting, presented in an influential real-time radiance-field formulation in 2023, represents a scene using many spatial Gaussian primitives with attributes controlling appearance and opacity. Projecting and blending these primitives can support fast novel-view rendering after optimization. Compared with neural fields evaluated through many network queries, it offers a different balance of memory, training behavior, rendering speed, and structural explicitness. Neither approach guarantees accurate hidden geometry or reliable views far outside the observations used to construct the scene.
Machine learning also supports denoising, image reconstruction, animation, material estimation, and content generation. Generative artificial intelligence, including diffusion models, can synthesize images or assist in creating textures and three-dimensional assets. A plausible image, however, is not necessarily a consistent scene model: different views may disagree about object shape, and visual realism does not establish geometric or physical accuracy. Evaluation therefore depends on the task. Film production may prioritize controllability and temporal consistency, while measurement applications require calibrated accuracy, uncertainty characterization, and a clear distinction between observed and inferred structure.
Production systems, scalability, and evaluation
A complete graphics application is a coordinated system rather than a rendering algorithm in isolation. Production pipelines connect modeling, materials, rigging, animation, simulation, lighting, rendering, and compositing. Each stage produces assets or derived data consumed by later stages. Scene organization must preserve object identity, hierarchy, units, coordinate conventions, and dependencies so that a change can propagate predictably. A revised character model, for example, may affect deformation, collision geometry, texture coordinates, and previously cached simulation.
Large scenes require selective representation. Level-of-detail methods replace detailed geometry or shading with cheaper alternatives when the difference is unlikely to matter in the final image. A distant building may use a simplified mesh, while a forest may combine instanced trees, impostor images, and procedural placement. Streaming loads resources as needed rather than keeping an entire environment resident in memory. These methods must account for transitions: abrupt changes in silhouette, shading, or density can remain noticeable even when the underlying objects are small.
Offline rendering often distributes frames or image regions across many machines using distributed computing. This introduces scheduling, storage, and reproducibility issues. A rendering task must receive the correct asset versions, renderer configuration, and random seeds, while failures must not invalidate completed work. Interactive systems face a different constraint: performance should remain stable under changing viewpoints and content. Average frame rate alone can conceal intermittent stalls caused by compilation, resource uploads, synchronization, or unexpectedly expensive views.
Quality evaluation combines objective measurement with perceptual judgment. Pixelwise differences are useful when a reliable reference image exists, but they can penalize visually harmless shifts or fail to describe important structural errors. Temporal stability, silhouette accuracy, material consistency, and task performance may matter more than a single-frame error score. Performance comparisons also require comparable conditions: resolution, sample count, scene complexity, hardware, reconstruction settings, and image quality targets can all change the apparent result. A renderer that is faster at one quality level may be slower at another.
Graphics production also intersects with intellectual property, provenance, and resource use. Assets may contain licensed geometry, textures, performances, or captured environments, making ownership and permitted reuse operational concerns. Images generated for documentation or scientific communication require traceable links to their underlying data and transformations. Computing cost depends not only on hardware power consumption but also on render duration, failed jobs, asset duplication, and the amount of iteration required. These constraints influence practical choices about scene complexity, simulation fidelity, sampling, and the division of work between artists, automated tools, and computing infrastructure.