Light, materials, and the rendered picture
Rendering turns a model into an image. One way is to simulate how light travels, and aim at a photographic result. Another is to apply an art style on purpose, which is what non-photorealistic rendering means. In a realistic renderer the two basic operations are transport, how much light gets from one place to another, and scattering, how a surface deals with the light that arrives. The work is done by 3D software or by a 3D graphics API. Getting the scene into a form the renderer can use also involves projection: a three-dimensional description displayed in two dimensions. Some modellers render as well. 3ds Max and Blender are examples. Some programs only render. Octane and Redshift are examples of that second kind.
Rasterization
Real-time 3D graphics usually forms the image by rasterization. Each polygon is projected into the frame, covers some pixels, and a depth test keeps the nearest surface. The question the machine asks is local and repetitive. Does this triangle cover this pixel, and is it in front of whatever was already there?
Gouraud shading computes a colour at the vertices and interpolates it across the face. Phong shading interpolates the normals and evaluates the highlight at each sample. The second is more expensive and, on a curved surface approximated by flat faces, the highlight sits more convincingly. Reyes, the architecture behind a lot of feature animation from the late 1980s on, dices each surface into micropolygons and then does its visibility test in screen space with a z-buffer. The dice is the point. If the pieces are small enough, a coarse description of a curved surface can be shaded as if the curve were really there.
Following the rays
Ray tracing forms the image by following rays from the camera through the scene. A primary ray leaves the camera through a pixel. In Whitted's model the shader may then spawn a reflection ray, a refraction ray, and a shadow ray toward each light. Recursion stops at a depth limit, or when a ray misses the scene entirely. The picture is whatever those rays found.
Path tracing estimates the rendering equation by averaging many random paths of light. Because each pixel is only a mean of random samples, stopping the average early leaves the image noisy. The noise is not a texture. It is the unfinished arithmetic, sitting in the frame where a later sample would have changed the mean.
By 2017 many production studios had moved from Reyes-style micropolygon rendering to physically based path tracing. Steve Bako and colleagues trained a convolutional network on frames from Finding Dory and tested it on Cars 3. From a few dozen samples per pixel, that network predicts a filter that removes a large part of the Monte Carlo noise. Intel's Open Image Denoise does related work, and renderers including Arnold, Blender, and V-Ray integrate it. In 2025 the Academy of Motion Picture Arts and Sciences gave a Technical Achievement Award to Attila T. Áfra for Open Image Denoise, and to Timo Aila for applying U-nets to denoising. The award is a useful marker. Cleaning up a render had become its own piece of film technology, not a checkbox.
Rays while you watch
Nvidia announced RTX on 19 March 2018, at first for Volta GPUs, together with Microsoft's DirectX Raytracing. Dedicated RT cores, for ray-traced shadows, reflections, refractions, and global illumination, came with the GeForce RTX series announced on 20 August 2018. Before a game could path-trace a whole frame, the usual practice was narrower. Shadows, or reflections, or indirect light, were ray traced as separate effects, and only for a limited number of lights. The rest of the frame was still rasterized.
Ray Tracing: Overdrive, added in Cyberpunk 2077 patch 1.62 on 11 April 2023, path traces the frame. It is off by default because the mode is very demanding. CD PROJEKT RED recommended that preview for GeForce RTX 40-series cards from the 4070 Ti upward, and for a GeForce RTX 3090 at 1080p and 30 frames per second. Those numbers will age. The shape of the fact will not. Full-frame path tracing in a game was, at that date, a high-end preview, not the default way a picture got onto a television.
Materials
Materials and textures are what the renderer consults when it has to decide how a surface treats light. A material is the instruction. A texture often supplies the numbers that vary across the surface. A colour map, also called an albedo map, paints the surface. A bump map or a normal map suggests smaller features than the mesh contains. A displacement map goes further and actually moves the surface. The first two are a lighting fiction. The third changes the shape that will be drawn, and it can change the silhouette, which a normal map cannot.
The Cook–Torrance model, from 1982, treats a surface as a collection of microfacets. It accounts for the relative brightness of materials shown together, for the directional spread of reflected light, and for a colour shift as the angle of incidence changes. Put a plastic cup next to a brass one under the same light and the model has something to say about both, rather than a single "shininess" number stretched over every substance.
Brent Burley's principled BRDF, described in a SIGGRAPH 2012 course, was built at Walt Disney Animation Studios around a short list of design rules. The parameters should be intuitive rather than strictly physical. There should be few of them. They should normally sit between zero and one, but be allowed past that range. They should stay plausible when you combine them. The studio used the resulting shader on virtually every material in Wreck-It Ralph. The interesting part of that list is the word "plausible." A film shader has to be controllable by an artist on Tuesday. A physically complete model that an artist cannot steer is the wrong tool, even if the integral is prettier.
Three later jobs, and one that is not the same
Denoising, reconstruction, and text-to-3D get bundled together because they all involve a network. They are different jobs. Denoising restores a picture the renderer was already simulating. Reconstruction fits a scene to photographs. Text-to-3D synthesizes a new asset from a written prompt. Off to the side of all three sits a generator that returns only a flat image and keeps no geometric model for a second viewpoint to share, even when the picture looks three-dimensional. That last kind is discussed with flat pictures. It does not belong in a file of models.
In denoising, a network writes a cleaned frame from the noisy render and, when the renderer exports them, from auxiliary buffers such as albedo and surface normals. What it estimates is the image a much larger sample count would have converged to. It is not inventing a scene. It is guessing the end of a calculation that was cut short.
Reconstruction starts from photographs whose camera poses are known. A neural radiance field, introduced in 2020 by Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi, and Ren Ng, maps a 3D point and a viewing direction to a density and a view-dependent colour. Along each camera ray, ordinary volume rendering samples that function and forms an image. The representation is accurate and expensive. Bernhard Kerbl and his colleagues, writing about what came next, describe neural radiance fields as costly to train and to render.
3D Gaussian splatting, published in 2023 by Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis, replaces the network with anisotropic 3D Gaussians. They are optimized, then drawn by splatting. On the scenes the authors tested, they report novel views in real time at 1080p, at or above 30 frames per second. The Gaussians are the stored scene. They are a radiance field drawn by splatting, not a mesh of triangles. In glTF, the KHR_gaussian_splatting extension can carry those splats, which is why they turn up in the list of model files.
DreamFusion, the text-to-3D method of Ben Poole, Ajay Jain, Jonathan Barron, and Ben Mildenhall, optimizes a randomly initialized neural radiance field until pictures from random cameras score well under a frozen text-to-image model. Because that prior was trained on ordinary images, the optimization does not need a corpus of 3D models. Each prompt is its own optimization. A known failure is the Janus artifact: more than one front face on a single object, which appears when each sampled view collapses toward a canonical front. Peihao Wang and colleagues trace that failure to a collapse in the score-distillation objective that DreamFusion made popular. The object is not being photographed from several sides. Each "side" is being tugged toward the front the image model knows best.
A production frame is still, most of the time, one of the older answers. Rasterize it if it has to move while someone watches. Trace paths if you can afford the wait, and then denoise. The newer representations matter when what you have is photographs, or a sentence, rather than a mesh someone built. From a model to a pixel is the vocabulary those choices are written in.