Updated on: 30 October 2026
Photon mapping is a two-pass global illumination algorithm that Henrik Wann Jensen developed between 1995 and 2001. In the first pass, the renderer shoots photons from the lights into the scene and stores where they land in a photon map; in the second pass, it traces rays from the camera and estimates the light at each visible point from the photons stored nearby. It is best known for caustics, the sharp patterns of light that glass and water focus onto other surfaces.
Most production renderers today are built on path tracing, but photon mapping is still in use, mainly for caustics. Below: how the two passes work, what the photon map stores, why the method is biased but consistent, how it compares with path tracing, progressive photon mapping, which render engines use photons today, and what AI image generators do and do not compute.
What Is Photon Mapping?
Jensen introduced the method in his 1996 paper "Global Illumination using Photon Maps" and described it in full in his book Realistic Image Synthesis Using Photon Mapping. Like every physically based method, it approximates the rendering equation (next section). Its target is the light that camera-based sampling alone finds slowly: caustics and the indirect light that bounces between surfaces, the core of global illumination.
Pass 1: photon tracing. Packets of light energy called photons are emitted from the light sources and traced through the scene. Where a photon hits a diffuse surface, the hit is stored in a spatial data structure, typically a KD-tree. That store is the photon map.
Pass 2: rendering. Rays are traced from the camera. At each visible point, the renderer estimates the light arriving there from the stored photons nearby (density estimation), instead of tracing every light path again at render time.
Because energy is traced from the light rather than found by random camera sampling alone, the method captures concentrated light well. Sunlight refracting through a glass façade, for example, can form sharp light patterns on a stone floor that a camera-side sampler finds only slowly. The price is extra memory and a precomputation step before the image.
Final Gathering
The photon map is rarely shown directly, because a direct lookup gives blotchy indirect light. Jensen's method splits the light at a visible point into four parts: direct light, computed with shadow rays to the lights; mirror-like reflection, computed with ray tracing; caustics, read straight from a dedicated caustic photon map; and the soft indirect light, estimated from the global photon map. In Jensen's implementation, that last part is computed with a final gathering step: from the shading point, many rays are sent into the scene and the photon map is looked up where they land, which averages out the unevenness of the photon density. On diffuse surfaces, irradiance caching can then reuse those results between neighbouring points.
The Rendering Equation and the Cost of Light Simulation
All physically based rendering techniques approximate the rendering equation, introduced to computer graphics independently by David Immel and colleagues and by James Kajiya in 1986. It describes the light leaving a point on a surface as the light the surface emits plus the incoming light it reflects, integrated over the hemisphere above the point.

Each term represents a physically meaningful quantity:
Lₒ is the outgoing radiance toward the viewer
Lₑ is the radiance the surface emits
Lᵢ is the incoming radiance from other directions
fᵣ is the bidirectional reflectance distribution function (BRDF), which models how the material responds to light
ωᵢ · n is the cosine term, which accounts for the angle between the incoming light and the surface normal
The equation is recursive: the incoming light at one point is the outgoing light of other points, which follow the same rule. Realistic rendering therefore becomes a high-dimensional integration problem with no closed-form analytical solution. Rendering engines approximate it with stochastic methods such as Monte Carlo sampling, the basis of path tracing, or with density-based techniques such as photon gathering.
The cost rises quickly because:
Every additional light bounce multiplies the number of possible light paths, especially with glossy or reflective materials.
Transparent materials need accurate refraction, which makes the paths longer and more complex.
Physically based rendering has to respect energy conservation, so every material response (BRDF) must be evaluated precisely.
Caustics are rare but concentrated energy paths that random sampling finds only slowly.
As a scene grows more complex, the number of possible light paths grows exponentially, and so does the computing time needed for clean, realistic light.
Russian Roulette in Photon Tracing
During photon tracing, each surface hit can reflect, transmit or absorb energy. Spawning a new photon for every possible event would make the photon count grow exponentially; after a few bounces it could no longer be computed.
Russian roulette controls this growth as a probabilistic termination strategy. Instead of splitting a photon into several branches, the algorithm picks one outcome at random, reflect, transmit or absorb, with probabilities given by the material's reflectance and transmission. Because the survival probability follows the material, a surviving photon keeps roughly the same power, and the stored energy is correct on average (in expectation).
This keeps the photon map filled with photons of comparable power and stops deep multi-bounce paths from multiplying. Russian roulette adds variance, but the result still converges as the number of photons grows, and photon tracing stays manageable.
What Gets Stored in the Photon Map?
Not every photon hit is stored. Photons are recorded only where they strike diffuse or non-specular surfaces. On a perfect mirror or ideal glass, a stored photon would carry no useful density information, so those events are handled with deterministic ray tracing in the rendering pass.
For each stored photon, the photon map records:
The surface position
The incoming photon direction
The photon's power (its flux)
A photon can be stored several times along its path if it hits several diffuse surfaces. Together, these records are the data the second pass uses for density-based radiance estimation: the photon map describes how light energy is spread over the diffuse surfaces of the scene, not the deterministic mirror reflections.
Biased vs Unbiased Estimators in Global Illumination
Photon mapping is a biased estimator: it estimates radiance from the density of photons within a finite search radius, so the result depends on the photon count and the gathering parameters. Averaging many photon-mapped renders of the same scene does not converge to the exact solution of the rendering equation.
It is, however, consistent: with more photons the render gets more accurate, and as the photon count approaches infinity, the result approaches the solution of the rendering equation.
Path tracing, by contrast, is an unbiased Monte Carlo estimator: on average (in expectation) it gives the correct answer, and with more samples it converges to the true solution. With a finite number of samples it shows variance, which appears as noise in the image.
In practice:
Biased methods such as photon mapping often reach a clean result faster for rare light events, at the cost of approximation error, usually blur, from density estimation.
Unbiased methods such as path tracing are correct on average, but at low sample counts their error shows as noise.
Both use variance reduction, such as importance sampling, stratified sampling and multiple importance sampling, to converge faster.
Photon Mapping and Caustics
Caustics are concentrated light patterns formed when light is reflected or refracted by specular materials and lands on a diffuse surface: the bright, focused patches that glass, water and polished metal cast. Typical examples in architecture:
Sunlight refracted by pool water, forming moving patterns on the pool floor
Glass façades and glass partitions projecting curved bands of light into a room
Crystal chandeliers and glass light fittings scattering sharp, focused highlights
These effects come from rare but visually dominant light paths. For a path tracer, which starts at the camera, they are hard to find: a ray has to bounce off the floor and happen to reach the small light source through the glass. Paths that are specular, then diffuse, then specular again, such as a caustic on a pool floor seen through the water surface or in a mirror, are harder still, and with a point light a camera-side sampler cannot find them at all.
Photon mapping solves this from the other side. It traces energy from the light source, through the glass or water, and stores where it lands; in the rendering pass, a dedicated caustic photon map gives the sharp pattern directly through local radiance estimation.
The trade-offs:
More memory for the stored photons
Bias that depends on the gathering radius and the photon count
Extra precomputation time before the final render
Blurred or blotchy caustics when the photon density is too low
Multiple Photon Maps: Global vs Caustic Maps
Practical implementations usually build more than one photon map, so that each part of the light is estimated with the photon density it needs.
A global photon map stores the photons that make up the diffuse interreflection across the whole scene, such as colour bleeding from a red wall onto a white ceiling. A separate caustic photon map stores only photons that passed through or bounced off a specular surface before reaching a diffuse one. Caustics need a much higher photon density to stay sharp, and this map is the only source of caustic light in the render, so it gets far more photons.
Extensions add a volume photon map for participating media such as smoke, fog or water vapour. Jensen's original method for subsurface scattering in translucent materials such as marble also used photon mapping; for highly scattering materials, BSSRDF models are now the more efficient choice.
Splitting the photons into specialised maps lets the rendering pass estimate each light component efficiently and with better control over quality.
When Should You Use Photon Mapping Today?
Path tracing dominates most production pipelines, but photon mapping still has an edge where the light is governed by rare, high-energy interactions, above all sharp caustics. Because it traces energy from the lights first, it reaches low-probability paths that camera-driven sampling resolves only slowly.
It is most useful when a scene includes:
Strong caustics from glass, water or crystal
Complex multi-bounce indirect light in enclosed spaces
Sharp specular-to-diffuse transitions, where a small set of paths carries much of the energy
Lighting in which finding the light paths is harder than reducing noise
In these cases photon mapping can give stable caustics with fewer camera samples than plain path tracing, as long as the extra memory and precomputation are acceptable. For large GPU renders with denoising, path tracing usually remains more scalable and simpler to integrate, which is why most engines use photons only as a caustics add-on, if at all.
So today photon mapping is mostly a tool for caustics, research and teaching, and a building block of hybrid methods that combine it with path tracing, rather than the default method for architectural visualization.
Which Render Engines Use Photon Mapping Today?
As of October 2026, the main render engines are path tracers, and photons appear mostly where caustics are concerned. What the engines' own pages say:
Engine | Caustics method | Photons involved? |
|---|---|---|
V-Ray (Chaos) | Photon-based caustics; quick caustics, added in V-Ray 7 update 3 for pools and water features | Yes in the photon-based setup; quick caustics avoid "the cost of full photon tracing" (Chaos) |
Corona Renderer (Chaos) | Fast Caustics Solver: reflective caustics for all materials, refractive caustics switched on per material | Yes: a combination of path tracing and photon mapping, a subset of vertex connection and merging |
OctaneRender (OTOY) | Photon Tracing kernel | Yes: GPU photon mapping combined with path guiding |
pbrt-v4 | SPPM integrator | Yes: stochastic progressive photon mapping |
Blender Cycles | Shadow caustics, based on MNEE | No photon map; only refractive caustics in the shadows of objects |
Mitsuba 3 | Path-based integrators | No photon mapping integrator in its list |
Sources, read in October 2026: Chaos's V-Ray 7 for 3ds Max what's new page, the paper on Corona's caustics solver by Martin Šik and Jaroslav Křivánek (EGSR 2019), OTOY's Photon Tracing kernel documentation, the pbrt-v4 file format, the Blender manual on object caustics and the Mitsuba 3 integrator list. How the two Chaos engines compare beyond caustics is covered in our Corona vs V-Ray guide.
Photon Mapping vs Path Tracing and Other Light Transport Methods
Path tracing remains dominant because it scales well and runs efficiently on GPUs. Several techniques extend it or combine it with photon mapping:
Bidirectional path tracing builds light paths from both the camera and the light sources, which raises the chance of capturing rare high-energy paths such as caustics.
Metropolis light transport uses Markov chain sampling to explore important light paths, concentrating effort where the image gets most of its light.
Vertex connection and merging (VCM), published in 2012 alongside the equivalent unified path sampling (UPS) of Toshiya Hachisuka, Jacopo Pantaleoni and Henrik Wann Jensen, combines bidirectional path tracing and photon mapping in one estimator, so each path is handled by the method that suits it. Corona's caustics solver is built on a subset of it.
Importance sampling and other variance reduction techniques steer samples toward the directions that carry the most energy, reducing noise without proportionally more samples.
Spectral rendering is a different kind of advance: it changes what is simulated, light per wavelength instead of three RGB channels, so that effects such as dispersion appear, rather than how fast an image converges. Photon mapping can be extended to it; see our guide to spectral rendering.
A simplified comparison:
Feature | Photon Mapping | Path Tracing |
|---|---|---|
Caustics | Efficient for concentrated paths such as refraction and focused reflections | Slow at low sample counts; needs many samples to resolve rare light paths |
Estimator type | Biased but consistent: density estimation over stored photons | Unbiased Monte Carlo estimator: correct on average, converges as samples grow |
Typical error | Blur or blotches when photons are too sparse | Noise (variance) at low sample counts |
Memory use | High: the photons are stored | Lower: relies on sampling rather than storage |
Convergence | Faster for rare energy events | Slower for rare paths, but no systematic error |
GPU scalability | Moderate; memory management is more complex | Strong; well suited to GPU rendering with denoising |
Progressive Photon Mapping and Modern Variants
Standard photon mapping uses a fixed photon count and a fixed gathering radius, so its bias depends on those parameters. Progressive photon mapping (PPM), published by Toshiya Hachisuka, Shinji Ogaki and Henrik Wann Jensen at SIGGRAPH Asia 2008, removes that limit with iterative refinement.
Instead of one large photon map, PPM starts with a ray tracing pass and then adds photon passes one after another. After each pass, the gathering radius shrinks while the number of photons grows, so the bias falls and the estimate converges to the correct result as the passes add up, without all the photons having to be kept in memory at once.
Later work extended the idea:
Stochastic progressive photon mapping (SPPM), Hachisuka and Jensen, SIGGRAPH Asia 2009: the progressive estimator for distribution effects such as depth of field, motion blur and glossy reflections
Parallel progressive photon mapping on GPUs, Hachisuka and Jensen, SIGGRAPH Asia 2010 Technical Sketches
Unified path sampling (UPS), 2012, which brings photon density estimation and bidirectional path tracing together and is implemented the same way as VCM
Hybrid methods that pair photons with path tracing, as today's caustics solvers do
The papers are listed on Toshiya Hachisuka's publications page, and pbrt-v4's SPPM integrator is an open implementation you can study.
Photon Mapping, PBR, and Energy Conservation
Physically based rendering rests on energy conservation: a surface cannot reflect more light than it receives. It uses physically plausible BRDF models, often based on microfacet theory, to describe how materials respond to light from different angles.
Photon mapping and path tracing both work inside this framework. They compute differently, but they approximate the same light transport with the same material models, so a correctly set up material behaves consistently in either. They are not separate philosophies, but two ways of solving the same physical equation.
Photon Mapping and AI Image Generators
AI image generators work in a completely different way. A diffusion model does not trace photons, store a photon map or solve the rendering equation: it generates a new image step by step from noise, guided by a prompt and, in architecture tools, by an uploaded picture. What it knows about light, how sunlight falls through a window or how glass shines, it learned from the images it was trained on. Our diffusion models guide explains the process.
That has two consequences. A generator can draw convincing light, reflections and even caustic-like patterns quickly, without photon storage, high Monte Carlo sample counts or recursive light evaluation. But nothing in the picture is computed: a caustic in an AI image is a plausible drawing, not a simulation, so it may break the physics of light and energy conservation.
Physically simulated data does play a role elsewhere in AI: neural denoisers and neural rendering research learn from path-traced images, and AI lighting models that predict light values are a field of their own. For the practical difference between computer-generated imagery and AI images, see how CGI and AI images differ.
That makes an AI tool useful before the physics, not instead of it. In ArchiVinci, you upload a screenshot of your model view, a sketch or a photo (JPG, PNG or WebP, up to 5 MB) and pick a lighting button such as Daylight, Golden hour or Twilight in the Exterior or Interior module: AI rendering generates a new image that follows your upload, in seconds, so you can compare light moods before you set up caustics in a physically based engine. The default engine, ArchiVinci 4.0, keeps the perspective and geometry of your view, so each light mood shows your own design.
Twilight gives dusk with the lights on; for a night version or any other light, you can relight a render in the Edit module by describing the light in one sentence. The light in these images is drawn, not simulated: for accurate caustics and lighting checks, the physically based renderer remains the tool.
Test light moods before the long render Upload a view of your model, a sketch or a photo and pick a lighting button such as Golden hour or Twilight: ArchiVinci generates a new image that follows your upload, in seconds. A new account starts with 3 free credits. |
Why Photon Mapping Matters in Architectural Visualization
Architecture is full of caustic-prone materials: glass façades and partitions, skylights, pools and water features, glass balustrades, crystal fittings and polished stone. In a final image, caustics are often the detail that makes such spaces read as real, and they are where a plain path tracer struggles most. That is why several engines used in archviz add a dedicated caustics solver, most of them built on photons (see the table above), instead of leaving caustics to plain path tracing.
In practice: switch caustics on only for the objects or materials that need them (Corona, for example, enables refractive caustics per material), test them in a small render region before the full image, and expect longer render times. For the wider topic of light in architectural images, see our guide to lighting rendering.
Early in a project, speed matters more than physical accuracy, and a quick AI image is often enough to choose a light mood. For the final image or a lighting check, a physically based renderer with path tracing, and photons where caustics matter, remains the tool.
Key Takeaways
Realistic lighting is expensive because it means approximating the recursive rendering equation.
Photon mapping (Henrik Wann Jensen, 1995–2001) traces photons from the lights, stores them in a photon map and reuses them in a second pass from the camera.
It is strongest at caustics and complex indirect light, the rare paths that random camera sampling finds slowly.
It is biased but consistent: more photons give a more accurate result. Path tracing is unbiased but noisy at low sample counts.
Global and caustic photon maps, final gathering and progressive photon mapping (PPM, SPPM) make it more stable and accurate.
As of October 2026, photons live on mainly in caustics: Corona's caustics solver, Octane's Photon Tracing kernel, V-Ray's photon-based caustics and pbrt-v4's SPPM.
AI image generators compute none of this: they draw light they learned from images, which is useful for quick light-mood studies but does not replace a physically based renderer.
Try it in ArchiVinci
Upload a sketch, a photo or a 3D model view and describe the result in a sentence: ArchiVinci generates a new image that follows your upload. A new account starts with 3 free credits.
Render a sketch, photo or model view: Render AI
From a photo or an image: AI image rendering
Frequently Asked Questions
Who invented photon mapping?
Henrik Wann Jensen developed photon mapping between 1995 and 2001. His 1996 paper "Global Illumination using Photon Maps" introduced it, and his book Realistic Image Synthesis Using Photon Mapping describes the method in full. Progressive photon mapping followed in 2008, by Toshiya Hachisuka, Shinji Ogaki and Jensen.
Is photon mapping still used?
Yes, mainly for caustics. As of October 2026, Corona Renderer's Fast Caustics Solver combines path tracing with photon mapping, OctaneRender has a Photon Tracing kernel, V-Ray offers photon-based caustics next to its newer quick caustics, and pbrt-v4 includes stochastic progressive photon mapping. As the main method for all the light in a scene, it has largely given way to path tracing.
What is the difference between photon mapping and path tracing?
Path tracing starts at the camera and follows random paths into the scene; it is unbiased, and its error shows as noise. Photon mapping first traces photons from the lights and stores them, then estimates the light from the stored photons; it is biased but consistent, and its error shows as blur. Photon mapping finds caustics far more easily, while path tracing scales better on GPUs.
Does Blender Cycles use photon mapping?
No. Cycles is a path tracer. As of October 2026, its shadow caustics option, based on manifold next event estimation (MNEE), speeds up refractive caustics in the shadows of objects you mark as caustic casters and receivers; it does not build a photon map.
What is the difference between radiance and irradiance in photon mapping?
Radiance is the light leaving a surface in a specific direction, while irradiance is the total light energy arriving at a surface from all directions. Photon density estimation starts from the incoming light the photons represent and, with the BRDF, turns it into the radiance seen by the camera.
Why are photons stored only on diffuse surfaces?
Density estimation needs a stable spatial distribution of light energy. Perfectly specular surfaces such as mirrors or ideal glass do not produce a meaningful density pattern, so those interactions are handled with deterministic ray tracing instead.
What is a KD-tree, and why does the photon map use one?
A KD-tree is a spatial data structure that organizes points in multi-dimensional space for fast nearest-neighbor and range searches. Photon lookups happen at every shading point, so their speed directly affects render time; Jensen advocates KD-trees because they find the nearest photons quickly even in large photon maps.
How does k-nearest neighbor search improve photon mapping accuracy?
Photon mapping often gathers the k nearest photons around a shading point. This adapts the effective search radius to the local photon density: small where photons are dense, larger where they are sparse, which improves stability and reduces estimation artifacts.
What causes blurring artifacts in photon mapping renders?
Blurring occurs when the gathering radius is too large or the photon density is too low. Because radiance is estimated as a density over an area, sparse photons smooth out sharp lighting features such as fine caustics. More photons, a smaller radius or the cone filter Jensen recommends, which weights each photon by its distance from the shading point, give sharper results.
Why is photon mapping considered a biased global illumination algorithm?
Because it estimates radiance from photon density over a finite radius, which introduces a systematic error. It is consistent, though: with more photons and a smaller gathering radius, the result approaches the solution of the rendering equation, which is what progressive photon mapping does automatically.
Do AI generators compute light the way photon mapping does?
No. A diffusion-based image generator does not trace light or solve the rendering equation; it generates an image whose light matches what it learned from its training images. Photon mapping explicitly simulates and records light interactions, while an AI generator only draws their visual result.
Can AI-generated images show accurate caustics?
They can show convincing caustic-like patterns, but those patterns are drawn, not computed: their shape and brightness need not be physically correct, and they do not follow energy conservation. For a caustic you have to rely on, such as sunlight through a glass façade onto a floor, use a physically based renderer with a caustics solver.
