On Graphics

---

Research checked: 20 September 2026. Scope: computer graphics broadly, including neural graphics and GPU computation.

Computer graphics turns a description of a scene into an image. That description might contain geometry, materials, lights, and a camera. The renderer determines what the camera would see. This framing remains useful as learned representations enter graphics: the central questions concern how a scene is represented, how an image is produced, and which kinds of control the representation allows.

Source: bicycle scene from Kerbl et al.’s 3D Gaussian Splatting project.

In traditional rendering, explicit geometry and material models provide a controllable world. Rasterization projects geometry into screen space, while ray tracing follows rays through a scene to determine visibility and light interactions. Physically based rendering uses models of light transport to reproduce effects such as reflection and indirect illumination. Real systems may combine techniques according to their performance and quality requirements. Pharr, Jakob, and Humphreys, Physically Based Rendering, fourth edition.

Neural graphics changes how some of that scene information is stored. NeRF represents a scene with a neural network that maps spatial position and viewing direction to density and color. Training uses multiple images with known camera poses, and volume rendering produces views from new camera positions. The original contribution demonstrated that a learned continuous representation could synthesize detailed novel views of a captured scene. Mildenhall et al., NeRF.

3D Gaussian Splatting uses a different representation: optimized three-dimensional Gaussian primitives with appearance and opacity information. Its rendering procedure projects and blends these primitives efficiently. The original work demonstrated real-time novel-view rendering at high visual quality on its evaluated scenes, showing that an explicit collection of learned primitives can offer a compelling speed-quality tradeoff. Kerbl et al., 3D Gaussian Splatting for Real-Time Radiance Field Rendering.

These methods solve a different problem from generating a single attractive picture from a prompt. Suppose a designer needs ten consistent views of a chair, with a precisely moved camera and a changed seat material. A representation that preserves geometry, camera relationships, and editable properties may be more useful than one image that looks excellent in isolation. This is a workflow judgment: the desired operations should determine which representation is appropriate.

The connection to AI hardware is also practical. Both graphics and neural models perform substantial parallel numerical work, but performance depends on moving data as well as doing arithmetic. FlashAttention illustrates this in machine learning: it reorganizes exact attention to reduce transfers between GPU memory levels. Its lesson extends to thinking about graphics workloads—an operation count alone does not describe the cost of executing a system. Dao et al., FlashAttention.

For a graphics project, I would evaluate realism together with editability, temporal consistency, rendering speed, and storage. A captured environment, a controllable game asset, and an illustration have different requirements. The interesting direction is the increasing range of combinations: explicit geometry for control, learned representations for reconstruction, and neural components for particular rendering tasks. The right system is the one that preserves the controls the work actually needs.