Physically based rendering (PBR) is an approach to computer graphics in which images are generated using mathematical models of light propagation, interaction with matter, and capture by an imaging system. Its defining feature is not a particular rendering algorithm or visual style, but an effort to make these models consistent with physical principles, including conservation of energy and the directional behavior of reflection and transmission. It supports effects such as soft shadows, metallic reflections, translucent skin, indirect illumination, and changes in material appearance under different lighting.
The term has two closely related uses. In research and offline image synthesis, it commonly describes a complete simulation framework encompassing illumination, materials, geometry, and cameras. In real-time production, “PBR” often refers more narrowly to a standardized material workflow based on quantities such as base color, roughness, and metallic character. Such a workflow may be combined with substantial approximations to light transport. Physically based rendering is therefore not synonymous with ray tracing, nor does it guarantee that an image is photorealistic or that its pixel values constitute a calibrated physical measurement.
The field combines optics, numerical integration, material science, and efficient computation. Its practical success depends on the interaction of these components: an accurate reflection model cannot compensate for incorrect lighting, and a sophisticated transport algorithm cannot recover geometric or material detail absent from the scene.
Physical quantities and the rendering equation
The basic language of physically based rendering is radiometry, the measurement of electromagnetic radiation in physical units. Radiant flux measures power in watts; irradiance measures incident power per unit area; and radiance measures power per projected area per solid angle. Radiance is particularly useful because, along an unobstructed ray in a nonparticipating medium with constant refractive index, it remains unchanged. This allows a renderer to connect a visible surface to a light source without introducing an inverse-square factor at every step along the ray.
For steady-state surface transport, the central relation is the rendering equation:
Here is outgoing radiance from position in direction , is emitted radiance, is incoming radiance, is the surface normal, and describes reflection. The integration covers the hemisphere above the surface. The cosine factor converts surface area to projected area as seen from the incoming direction. Transmission requires additional terms or a scattering function defined on both sides of the surface.
This equation is recursive: incoming light at one surface can be outgoing light from another. A red wall can therefore illuminate a nearby white object with reddish light even when neither surface emits energy. Such interactions are part of global illumination, which includes interreflection, indirect shadows, and light paths involving multiple surfaces. Direct illumination considers light arriving from emitters without an intervening scattering event; indirect illumination includes one or more such events.
Inverse-square behavior appears when converting between descriptions of illumination. An ideal isotropic point source with radiant intensity produces irradiance on a small receiving surface at distance . An extended source behaves differently at short range because it occupies a substantial solid angle. Its apparent brightness and its total illumination must not be confused: a uniformly emitting panel can retain approximately constant radiance while contributing less irradiance as it recedes.
Most production rendering assumes steady illumination, geometrical optics, and incoherent light. These assumptions exclude or simplify interference, diffraction, polarization, and time-dependent propagation. They are effective for many scenes whose important structures are large relative to visible wavelengths. However, thin films, diffraction gratings, strongly polarized reflections, and ultrafast imaging may require extensions beyond the conventional surface rendering equation.
Surface scattering and material models
A material model specifies how incident light is reflected, transmitted, or absorbed. The bidirectional reflectance distribution function (BRDF) relates incident irradiance from one direction to reflected radiance in another. Its units are inverse steradians. For an ordinary passive reflecting surface, its hemispherical integral cannot exceed the incident energy:
For reciprocal materials under the usual assumptions, exchanging incident and outgoing directions leaves the BRDF unchanged. Reciprocity is important both as a physical constraint and as a foundation for rendering algorithms that trace light paths in opposite directions.
An ideal Lambertian surface has , where is its diffuse reflectance. This model produces outgoing radiance independent of viewing direction for fixed incident illumination. It is useful for approximating matte surfaces, but real diffuse reflection often results from light entering a material, scattering internally, and returning through the surface. Rough plaster, fabrics, and powders can depart substantially from Lambertian behavior, especially near grazing angles.
Glossy reflection is commonly represented by microfacet theory. A rough surface is treated statistically as a collection of small reflecting facets. A typical reflection model has the form
where is the half-vector between incident and outgoing directions. The distribution describes facet orientations, describes angle-dependent reflection, and accounts for masking and shadowing among facets. Beckmann and Trowbridge–Reitz, often called GGX, distributions produce different highlight shapes and tails. Anisotropic distributions allow different roughness along two tangent directions, as in brushed metal.
The Fresnel equations describe reflection and transmission at an interface. Dielectrics such as glass and clear plastic generally have weak normal-incidence reflection that increases toward grazing angles. At normal incidence, the reflectance of an ideal interface between nonabsorbing media is
For air and glass with refractive indices approximately 1 and 1.5, this gives about 4 percent. Conductors require a complex refractive index and commonly produce wavelength-dependent, colored reflection. Bulk metals do not have the same diffuse component as pigmented dielectrics, although coatings, oxides, and contamination can introduce additional behavior.
Single-scattering microfacet models can lose too much energy at high roughness because they omit light reflected between facets. Multiple-scattering models or energy-compensation approximations address this limitation. Layered materials add further complexity: a clear coat can reflect some light before it reaches an underlying paint layer, while the remaining light crosses the interface again on exit. Simply adding independent lobes without accounting for these exchanges can produce excessive reflected energy.
Geometry, textures, and participating media
Physical appearance depends on geometry over many scales. A curved metal object, a scratched metal sheet, and a statistically rough metal surface may share optical constants but scatter light differently because their structures occur at different sizes. Renderers distribute these structures among explicit geometry, displacement, surface-normal perturbations, and statistical scattering models. The division is partly computational: a feature visible in silhouette usually requires geometry, whereas unresolved scratches may be represented through an anisotropic reflection model.
Scenes commonly use a polygon mesh to approximate surface shape. Texture mapping assigns spatially varying material parameters through coordinates attached to that surface. A texture can encode base reflectance, roughness, opacity, displacement, or other quantities; it is not inherently a photograph. A photograph used directly as base color may contain shadows and highlights that the renderer then generates a second time, producing lighting that is effectively baked into the material.
Normal mapping changes the normal used for shading without changing the actual surface. It can create convincing small-scale relief but cannot, by itself, alter silhouettes or reproduce all self-occlusion. The geometric normal and shading normal consequently serve different purposes. Poor handling of their difference can cause light leaks, dark boundaries, or inconsistencies between forward and reverse transport. Filtering is equally important: a distant textured surface covers many texels per pixel, and naïve sampling produces flicker or aliasing. Mipmapping stores progressively filtered texture levels, although filtering roughness or normal distributions correctly is more involved than averaging ordinary color.
Some materials cannot be modeled adequately by reflection at a single surface point. Subsurface scattering describes light entering an object, scattering internally, and emerging elsewhere. It is important for wax, marble, milk, leaves, and skin. A bidirectional scattering-surface reflectance distribution function relates illumination at one position to outgoing light at another. Diffusion approximations are efficient when scattering is strong and the material is sufficiently thick, but become less reliable near thin structures, boundaries, and sources. Explicit random walks can represent a broader range of internal paths at greater computational cost.
Fog, smoke, clouds, and water containing suspended particles require transport through a volume. The radiative transfer equation accounts for extinction, emission, and scattering into a direction along a ray. In a homogeneous medium, unscattered transmission over distance follows
where is the extinction coefficient. A phase function describes the angular distribution of scattered light; cloud droplets, for example, scatter strongly in forward directions. Multiple volume scattering creates soft illumination and can dominate computational cost. Surface roughness, spatial texture, and volumetric scattering thus represent different mechanisms rather than interchangeable ways of making an object appear less sharp.
Numerical light transport and sampling
The rendering equation rarely has a closed-form solution for a realistic scene. Rendering algorithms therefore approximate its integrals over directions, positions, wavelengths, lens coordinates, and time. Monte Carlo integration estimates an integral by evaluating samples drawn from a probability density :
Under suitable conditions this estimator is unbiased, meaning its expected value equals the desired integral. For independent samples with finite variance, its standard error decreases proportionally to . Reducing noise by a factor of two therefore ordinarily requires about four times as many samples.
Path tracing constructs light paths by repeatedly sampling scattering events, commonly starting at the camera. At each intersection, it evaluates emission, estimates direct lighting, and samples a continuation direction. The contribution is weighted by scattering factors and sampling probabilities. Russian roulette can terminate long paths without bias by probabilistically discarding them and appropriately increasing the weight of survivors. A fixed maximum path depth, by contrast, generally introduces truncation bias, although it may be useful for controlling cost.
Good sampling is essential because visible light often arrives through a small part of the integration domain. Importance sampling concentrates samples where contributions are expected to be large. Sampling a small light source directly is usually much more effective than hoping a randomly reflected ray strikes it. Conversely, sampling the material response is important for a sharply reflecting surface. Multiple importance sampling combines strategies with weights that reduce the penalties when one strategy performs poorly.
Different transport algorithms handle different path structures. Bidirectional path tracing builds subpaths from both the camera and the lights and connects them. Metropolis light transport mutates existing paths, concentrating work on important regions of path space. Photon mapping stores light-transport samples and reconstructs illumination from nearby samples; finite-radius estimates introduce bias, while progressive variants can reduce that bias as computation increases. Focused patterns formed by reflection or refraction, called caustics, are particularly difficult for ordinary camera-path tracing when their contributing paths are improbable.
Ray intersection is accelerated using a bounding volume hierarchy or another spatial data structure so that each ray need not test every primitive. Numerical robustness is also necessary: rounding errors can cause a newly spawned ray to strike the surface it just left, creating false shadows or other artifacts. Parallel computing distributes samples across processors, while scheduling and memory layout affect how efficiently hardware executes divergent paths. Denoising can remove visible noise using neighboring pixels, temporal history, or learned priors, but it changes the estimator and may erase small features or reconstruct details not supported by enough samples.
Real-time rendering and practical PBR workflows
Real-time rendering must deliver images within a strict frame budget. At 60 frames per second, the entire frame has approximately 16.7 milliseconds, including simulation, visibility, lighting, and post-processing. This makes an exhaustive solution of light transport impractical for many interactive scenes. Real-time PBR instead combines physically motivated material models with approximations selected for stability, speed, and predictable authoring.
A common material representation is the metallic–roughness workflow. Base color represents diffuse reflectance for a dielectric and approximately the normal-incidence reflection color for a metal. A metallic parameter chooses or blends between those cases, while roughness controls the spread of glossy reflection. Intermediate metallic values can represent filtered mixtures or transitional texels rather than a homogeneous substance that is literally partly metallic. Alternative specular–glossiness workflows expose diffuse and specular colors more directly. Neither representation describes every physical material, and converting between them can lose information.
An artist-facing roughness value is not a universal measured quantity. One renderer may square it before passing it to the microfacet distribution, while another may use a different mapping. Normal-map orientation, tangent-space construction, transmission conventions, and clear-coat behavior can also vary. Matching texture names across two applications therefore does not guarantee matching appearance. Consistency requires agreement on the underlying equations, parameter transforms, texture interpretation, and lighting.
Rasterization efficiently determines which triangles cover screen pixels, after which a shader evaluates the material and illumination. Deferred shading stores surface attributes in intermediate buffers and applies lights later; forward approaches evaluate lighting during surface shading. These architectures concern visibility and execution rather than the physical validity of the scattering model. Either can use energy-aware materials, and either can be combined with ray-traced effects.
Image-based lighting represents surrounding illumination with an environment map. Preconvolving that map for diffuse or rough specular responses replaces expensive directional integration with texture lookups. A common specular approximation separates environment filtering from a precomputed BRDF integration term. This is efficient but imperfect, especially when local geometry should occlude or distort reflections. Reflection probes, baked light maps, shadow maps, and screen-space effects supply other pieces of the transport solution.
Modern graphics processing units can accelerate ray traversal and intersection, enabling hybrid pipelines with ray-traced shadows, reflections, or indirect illumination. Low sample counts make spatial and temporal reconstruction important. Reusing previous frames can greatly improve apparent quality, but moving objects, disocclusion, and changing light sources can produce ghosting or lag. These systems are physically based in their modeled quantities and scattering laws, yet their final images remain products of approximated transport, reuse, and reconstruction.
Spectral rendering, cameras, and color pipelines
Light is distributed over wavelength, whereas most images store only three color channels. Conventional RGB rendering performs transport using three-component values associated with a chosen color space. This is computationally efficient and often visually effective, but RGB values are not complete descriptions of spectra. Two different spectra can produce the same color response for one observer or illuminant and differ under another, a phenomenon known as metamerism.
Spectral rendering samples or discretizes wavelength and evaluates wavelength-dependent emission, scattering, and absorption. It can represent dispersion in glass, strongly colored absorption, narrow-band lighting, and material behavior that RGB approximations handle poorly. A dispersive prism, for example, requires the refractive index to vary with wavelength. Spectral rendering nevertheless needs suitable spectral input data; converting an RGB texture into a spectrum is underdetermined because many spectra correspond to the same three values.
A physically based camera model integrates radiance over the image footprint, lens aperture, shutter interval, and spectral sensitivity. A pinhole approximation produces ideal perspective but no finite-aperture depth of field. A thin-lens model introduces a focal plane and circles of confusion, while more detailed lens models trace rays through multiple optical elements. Sampling time within the shutter interval produces motion blur. Rolling shutters, lens distortion, vignetting, and wavelength-dependent aberrations require additional camera-specific treatment.
The relationship between exposure and recorded image values must be distinguished from the display transform. Scene-linear values are proportional to modeled light before nonlinear encoding or artistic mapping. Doubling exposure doubles these linear values under an ideal unclipped response. An actual sensor additionally exhibits effects such as photon shot noise, electronic read noise, saturation, and spatial variation. A renderer may simulate these effects for camera matching, but they are not inherent requirements of every PBR image.
Tone mapping compresses or reshapes scene values for a display with limited luminance and color gamut. A bright window and a dim room may coexist in the simulation even when a particular display cannot reproduce their physical luminance ratio. Exposure, tone curves, gamut mapping, and viewing conditions determine how that range becomes a visible image. Thus a final display pixel is not generally a direct radiance measurement.
Texture decoding is a frequent source of error. Color textures stored with an sRGB transfer function must normally be decoded before linear lighting calculations. Roughness, metallic masks, and normal maps represent data rather than display colors and usually must not receive that same decoding. Similarly, adding light contributions in gamma-encoded values produces incorrect results. A consistent pipeline keeps physical calculations, storage encodings, artistic grading, and display presentation conceptually separate.
Historical development
Physically based rendering emerged from several lines of research rather than a single invention. Early computer graphics often prioritized recognizable images and efficient display, using local shading formulas that were only loosely connected to optics. In parallel, research in illumination engineering, heat transfer, and optical scattering supplied more rigorous models. Torrance and Sparrow’s 1967 treatment of off-specular reflection helped establish the statistical rough-surface ideas that later became central to computer graphics materials.
Turner Whitted’s 1980 paper on an improved illumination model demonstrated recursive ray tracing for mirror reflection, refraction, and shadows. Its importance lay in following structured chains of visibility through a scene, but the method did not generally integrate diffuse interreflection. Robert Cook and Kenneth Torrance introduced their influential reflectance model for graphics in 1982, connecting material appearance to microfacet orientation and Fresnel behavior. Cook, Thomas Porter, and Loren Carpenter’s 1984 distributed ray tracing work used sampling to represent effects including soft shadows, motion blur, and depth of field.
In 1986, James Kajiya presented the rendering equation and a path-tracing approach to its numerical solution. In the same year, David Immel, Michael Cohen, and Donald Greenberg presented an equation-based treatment that generalized radiosity beyond ideal diffuse reflection. Radiosity methods were especially influential for architectural interiors because they represented diffuse energy exchange between surfaces, producing effects such as soft indirect lighting and color bleeding. Their discretization and view-independent solutions differed from the camera-driven sampling strategies that later became dominant.
During the 1990s, research broadened the range of efficiently sampled light paths. Eric Veach and Leonidas Guibas introduced multiple importance sampling in 1995 and Metropolis light transport in 1997. Henrik Wann Jensen introduced photon mapping in the 1990s, providing an influential approach to caustics and indirect illumination. These methods addressed a central difficulty of physically based simulation: a correct mathematical model can still be computationally ineffective when important paths are rarely sampled.
The first edition of Physically Based Rendering: From Theory to Implementation, by Matt Pharr and Greg Humphreys, appeared in 2004. Its accompanying renderer, pbrt, provided an influential, openly inspectable integration of transport theory and implementation. Production adoption also depended on more practical material controls. Brent Burley’s 2012 presentation of Disney’s principled shading model emphasized an artist-oriented parameterization grounded in measured material behavior. Brian Karis’s 2013 account of shading in Unreal Engine 4 helped document the adaptation of similar principles to real-time constraints. The material-workflow meaning of “PBR” became especially prominent during this period, alongside continuing advances in full light-transport simulation.
Validation, uncertainty, and limits of physical accuracy
A renderer can be mathematically consistent with its implemented equations and still disagree with a photograph. Validation therefore has several levels. Individual components can be checked against analytical results, such as Lambertian reflection under uniform illumination or Fresnel reflectance at a planar interface. More complex transport can be compared with trusted numerical solutions. End-to-end validation requires a scene whose geometry, illumination, materials, camera response, and processing are sufficiently characterized to support a meaningful comparison.
Energy tests are useful for material models. A white-furnace test surrounds a surface with uniform radiance and examines whether its reflected response obeys the intended energy balance. Reciprocity tests exchange incident and outgoing directions. These checks can reveal normalization errors, incorrect probability densities, or energy losses in a model. They do not establish that the model matches a particular real material: a reciprocal, energy-conserving BRDF can still have the wrong angular appearance.
Measured appearance data also contain limitations. A reflectance instrument samples a finite set of directions and wavelengths, with errors in alignment, illumination, and detector response. Very narrow specular peaks are difficult to capture, and fitting a compact analytic model may smooth away important structure. Surface contamination, wear, moisture, and manufacturing differences can make two nominally identical materials look different. Consequently, assigning a library material called “aluminum” or “oak” does not uniquely specify a physical surface.
Scene scale and illumination units affect validation. A light defined in watts, lumens, or candela cannot be substituted without accounting for what the unit measures; lumens and candela incorporate human visual sensitivity, unlike purely radiometric power. Incorrect geometric scale changes distances, volumetric attenuation, and subsurface path lengths. Camera exposure and nonlinear image processing can conceal or exaggerate discrepancies, so a visually similar pair of display images need not imply matching radiance.
Numerical error and modeling error must also be separated. More samples can reduce Monte Carlo noise but cannot correct an inappropriate scattering law, missing geometry, or an inaccurate light spectrum. A denoised image may appear cleaner while deviating more from the true solution in small details. Finite path depth, clamping unusually bright samples, and approximate visibility introduce further tradeoffs.
Some phenomena require models outside standard production PBR. Coherent interference, detailed polarization, fluorescence, and nonlinear optical effects need additional state or different transport assumptions. Fluorescence, for instance, transfers absorbed energy from one wavelength range to another rather than treating each wavelength independently. The appropriate level of physical detail depends on which measured or perceptual effects the image is intended to reproduce.
Applications and inverse problems
Film and visual-effects production use physically based rendering to make synthetic objects respond coherently to complex lighting and to integrate them with photographed environments. Measured illumination, reconstructed cameras, and matched exposure help place a digital object within a live-action image. Physically based shading also reduces the need to invent separate material settings for every shot, although production lighting, compositing, and material adjustments remain artistic processes rather than purely predictive measurements.
Product visualization uses the same consistency to compare finishes, shapes, and lighting arrangements before physical prototypes are available. Automotive paint, coated glass, textiles, and polished components are demanding cases because their appearance depends strongly on angle and material structure. Architectural visualization similarly benefits from extended light sources, indirect illumination, and physically meaningful scale. However, a visually persuasive architectural image is not automatically a validated daylighting calculation: quantitative analysis requires appropriate sky models, material measurements, and verified outputs.
Interactive applications, including games and virtual reality, benefit from reusable assets whose appearance remains coherent as the viewer and lights move. The rendering budget constrains the accuracy of indirect illumination, reflections, and transparency. In virtual reality, low latency and consistency between the two eyes impose further requirements; unstable reconstruction or mismatched effects can remain conspicuous even when individual frames look convincing.
Physically based rendering also supports computer vision by generating synthetic images with known scene properties. A renderer can provide depth, surface normals, object identities, and motion alongside color. Such images can augment training data for perception systems, but realism alone does not guarantee transfer to real environments. Differences in object distributions, sensor behavior, backgrounds, and lighting can create distribution shift. Controlled variation and camera-specific simulation may matter as much as the sophistication of the reflection model.
Inverse rendering reverses the usual process: instead of generating an image from a scene, it estimates scene properties from images. Unknowns may include shape, reflectance, illumination, or camera parameters. The problem is often ambiguous because a dark pixel could indicate dark material, weak illumination, shadow, or an unfavorable viewing angle. Multiple views, known lights, measured reference objects, and physical constraints can reduce these ambiguities without eliminating them.
Differentiable rendering supplies derivatives of an image or objective with respect to scene parameters, enabling gradient-based optimization. Smooth material changes are comparatively straightforward, while visibility changes at silhouettes and shadow boundaries require special treatment. Applications include material capture, geometry reconstruction, and fitting digital scenes to observations. The recovered parameters remain tied to the chosen forward model: if that model omits a coating, scattering mechanism, or camera response, optimization may compensate by assigning incorrect values to other parameters.