Photorealism in digital imagery developed through several related but distinct fields: computer graphics, physically based rendering, ray tracing, image-based rendering, neural scene representations and generative AI. These methods do not simply replace one another. They solve different problems and often appear together in production pipelines.
For a technical account of NeRFs and 3D Gaussian Splatting, see our guide to neural rendering technology.
Early computer graphics and rendered imagery
Early computer-graphics research established geometric modelling, shading and lighting methods. The Utah teapot became a widely used test model in the 1970s. Feature films later demonstrated increasingly sophisticated CGI, while physically based rendering and ray-tracing research improved the simulation of light transport, reflections, shadows and materials.
Physically based rendering
Physically based rendering uses models of material response and illumination to produce consistent results across lighting conditions. Path tracing and related ray-tracing techniques estimate light transport through repeated sampling. Their limitations include computational cost, noise at low sample counts and the need for accurate scene geometry and materials.
Neural rendering and scene reconstruction
Neural Radiance Fields (NeRFs) represented scenes through a learned function for density and view-dependent colour. The original NeRF research was published in 2020. 3D Gaussian Splatting, published in 2023, introduced a different explicit scene representation designed for real-time radiance-field rendering. See the NeRF research project and 3D Gaussian Splatting.
Generative image models
Generative models create images from learned data distributions. GANs demonstrated image synthesis through adversarial training; diffusion models later became prominent for text-to-image generation and editing. These systems can create plausible images without a fully modelled 3D scene. Their outputs may not preserve physical consistency, exact geometry or factual accuracy.
For a related discussion of open-weight generative models and deployment choices, see open-weight generative AI models.
Hybrid production pipelines
Modern workflows may combine conventional 3D assets, ray tracing, neural scene representations and generative image tools. A virtual-production scene may use a real-time renderer for the environment, tracked camera data for perspective and post-production tools for finishing. A concept-art workflow may use generative imagery as a starting point and then move into a conventional 3D or compositing pipeline.
Evaluate photorealism
Assess lighting consistency, geometry, material response, reflections, shadows, temporal stability, detail and editability. Use a defined reference scene and record model/software versions. A visually convincing still image is not necessarily physically accurate or production-ready.
For the relationship between computational capture and image reconstruction, see computational visuals and their limits.
Conclusion
Photorealism evolved through CGI, ray tracing, image-based methods, neural rendering and generative AI. The methods are complementary and differ in control, editability, computational cost and factual reliability. A credible history names the milestones and explains their distinct roles.