When computer pictures became ordinary
The obvious change in the 1990s was that three-dimensional modelling got cheap enough to be common, and that CGI in general got better in a way a non-specialist could see. A home computer could take on rendering jobs that had belonged to workstations costing thousands of dollars. Once modellers ran on those home machines, the Silicon Graphics workstation lost the automatic place it had held, and Windows PCs and Macintoshes running Autodesk tools such as 3D Studio, or other rendering software aimed at a desk, became the ordinary choice. By the end of the decade the graphics chip was on the way to the position it still holds.
For the first time, some rendered pictures could pass as photographs to an untrained eye. A trained CGI artist was not fooled, and the gap between those two viewers is the plot of the next twenty years. 3D graphics also became ordinary in games, in multimedia, and in animation, rather than a special sequence you pointed at.
Television, then a toy story
At the end of the 1980s and the start of the 1990s, France produced the first computer-graphics television series. La Vie des bêtes, from Mac Guff Ligne, dates to 1988. Les Fables géométriques, from studio Fantôme, ran from 1989 to 1991. Quarxs, by Maurice Benayoun and François Schuiten at Z-A, ran from 1990 to 1993 and was the first computer-graphics series made for high-definition television. Series matter more than a one-off film. A series means the pipeline can be run again next week.
In film, Pixar's commercial rise belongs to this decade. Toy Story, in 1995, directed by John Lasseter, was a critical and a commercial success on a scale of nine figures. The models were built on Silicon Graphics and Sun workstations. RenderMan produced the frames. It was the first theatrical feature whose imagery was entirely computer-generated. The studio that had invented the programmable shader kept making animated films after that, and its prerendered work stayed a reference point for everyone else. Edwin Catmull's old goal, a feature made this way, had a release date.
Games that were actually 3D
In 1992 Virtua Racing, on Sega's Model 1 arcade board, laid the ground for fully 3D racing games and showed real-time polygonal graphics to a wide game audience. The Model 2 in 1993 and the Model 3 in 1996 pushed that hardware further. On the PC, id Software released Wolfenstein 3D, Doom, and Quake across the decade, using a rendering approach developed chiefly by John Carmack. They were enormously popular, and they taught a generation of players what a first-person view computed in real time felt like. The distinction is worth keeping. Doom's world is not the same kind of 3D as a later game with an unrestricted camera. Players experienced it as depth. The restriction was a technical choice that also became a style.
The Sony PlayStation, the Sega Saturn, and the Nintendo 64 sold in the millions and put 3D in the living room. A handful of late-1990s games did a disproportionate amount of the teaching: Super Mario 64, The Legend of Zelda: Ocarina of Time, and fighting games such as Virtua Fighter, Battle Arena Toshinden, and Tekken. After those, a console that could not do 3D looked unfinished, whether or not 3D was what a particular game needed.
A chip that called itself a GPU
The algorithms kept moving. In 1996 Venkat Krishnamurthy and Marc Levoy described normal mapping, an improvement on Jim Blinn's bump mapping. A normal map stores a direction rather than a height, and the lighting then behaves as if the surface were more detailed than the mesh. In 1999 Nvidia released the GeForce 256, the first home video card marketed as a graphics processing unit. In the company's own words it contained "integrated transform, lighting, triangle setup/clipping, and rendering engines." One chip did the geometry and the pixels. Nvidia called it the world's first GPU.
That claim needs the earlier chip beside it. NEC's μPD7220, from the early 1980s, was already a single integrated graphics processor, the first of its kind. GeForce 256 is a first at a different job: transform and lighting on the card, in a product sold for a home PC. By the end of the 1990s, machines had also settled on shared programming interfaces, DirectX and OpenGL. AMD, through ATI, became the other large maker of graphics boards. The duopoly dates from this decade.
The 2000s, and a valley in the graph
By the late 1990s, games and CGI cinema had carried computer graphics into the mainstream. The 2000s accelerated that. CGI was adopted in bulk for television advertising, so people who never played a game and never sought out an animated film still saw the pictures every evening. The graphics chip kept getting more capable, and a 3D-capable GPU became something a desktop maker was expected to include. Nvidia's GeForce line dominated the early part of the decade, with ATI as the serious competitor. Later, even cheap machines usually had some sort of 3D chip, as both Nvidia and AMD sold low-priced parts and kept the market between them.
Shaders, introduced in the 1980s as specialised programs on the graphics hardware, were supported on most consumer cards by the end of the 2000s. That sped the pictures up and made detailed texture and shading normal rather than exceptional. Normal mapping, bump mapping, and a family of related tricks let a relatively coarse mesh read as a detailed surface. The pictures in films and games got close enough to photographic that they walked into the uncanny valley: the dip in comfort that is said to appear when a human face is almost right and still wrong.
Final Fantasy: The Spirits Within, in 2001, was the first fully computer-generated feature to use photorealistic characters and to be made entirely with motion capture. It was not a box-office success. Some commentators blamed the faces, which sat in that valley. The Polar Express drew a similar kind of attention. Other CGI films simply declined the bet. Ice Age, Madagascar, and Pixar films such as Finding Nemo dominated the animated box office as cartoons that were three-dimensional in their models and plainly not trying to be mistaken for a photograph. Star Wars returned as a prequel trilogy and kept resetting what a live-action film was expected to spend on digital pictures.
On consoles, the PlayStation 2 and 3, the Xbox line, and Nintendo's GameCube held large audiences, and so did the Windows PC. Series such as Grand Theft Auto, Assassin's Creed, Final Fantasy, BioShock, Kingdom Hearts, and Mirror's Edge pushed toward photographic detail and grew the game business until its revenues were spoken of in the same breath as cinema. Microsoft tried to make DirectX easier for independent developers with XNA. The program was not a success. DirectX itself was. OpenGL matured alongside it, and the second-generation shader languages HLSL and GLSL became the normal way to write the interesting part of a renderer.
In scientific computing, GPGPU appeared: moving large amounts of data both ways between a graphics chip and a CPU, so that the graphics chip could be used as a calculation engine. It sped up work in bioinformatics and molecular biology. The same idea was turned to Bitcoin mining, and it has uses in computer vision. A chip designed to colour triangles had become a general arithmetic machine that happened to be good at triangles.
The 2010s and after
In the 2010s, CGI was close to ubiquitous in video. Prerendered graphics could be photorealistic in a scientific sense, and real-time graphics on a high-end system could pass for that, to an untrained eye, for a while. Texture mapping had become a stack of stages. It is normal to combine a colour texture, a bump or normal map, lighting maps, reflection, and shadow volumes in one engine, driven by shaders. Shaders are close to mandatory for advanced work. They let a program touch pixels, vertices, and textures one element at a time, which is an enormous amount of rope. HLSL and GLSL stayed active research subjects. So did physically based rendering, which uses many maps and a heavier lighting calculation so that materials behave more like optical surfaces than like paint. Ambient occlusion, subsurface scattering, Rayleigh scattering, photon mapping, and ray tracing all sat in the same busy margin. Experiments in real-time 4K began. They were out of reach for anything but the top hardware.
In cinemas, most animated features were CGI, and few of them chased photorealism. The fear of the uncanny valley had not gone away. Most of what got made was a three-dimensional cartoon. In the living room the Xbox One, the PlayStation 4, and the Nintendo Switch were the home consoles, all of them capable of advanced 3D, and Windows remained one of the main gaming platforms.
In the 2020s, ray tracing became usable while a picture was still on the screen. Waiting overnight was no longer the only way to see it. AI methods arrived for generating frames or for upscaling them. Nvidia pushed both, with ray-tracing cores and with DLSS and tensor cores. AMD followed, with FSR and with its own ray-tracing hardware. The argument that started with Appel's rays in 1968, about whether you follow the light or you fake it fast enough, is now an option in a graphics card menu. A method that once required the most powerful computer on earth became a setting.