Rendering is the computational process of generating a finished visual image or video frame from underlying data: converting abstract numerical, geometric, or probabilistic information into the pixel values of a viewable image. In traditional computer graphics and 3D production, rendering refers to the calculation of how light interacts with surfaces, materials, and camera optics within a three-dimensional scene model to produce a photorealistic or stylised two-dimensional output. In AI generation contexts, rendering refers more broadly to the inference process through which a model transforms its internal representations: latent vectors, noise fields, weight-conditioned activations: into the final visible image or video that the user receives. In both senses, rendering is the final computational step that converts data into image, and its quality: the fidelity, detail, coherence, and visual accuracy of the result: is the ultimate measure of the entire upstream process.
In traditional 3D computer graphics, rendering is the most computationally intensive stage of production, requiring the simulation of physical light behaviour at a level of detail sufficient to produce convincing imagery. Ray tracing, path tracing, rasterisation, and radiosity are the primary rendering algorithms, each making different trade-offs between computational cost and physical accuracy. Ray tracing and path tracing simulate light rays bouncing through a scene, accurately computing shadows, reflections, refractions, and global illumination at high computational cost. Rasterisation projects 3D geometry onto the 2D image plane directly, producing results quickly enough for real-time applications but with less physically accurate lighting. The distinction between real-time rendering (video games, interactive applications) and offline rendering (film VFX, architectural visualisation, animation) reflects the different computational budgets available: offline renders for major film productions may take hours per frame on large server farms, while real-time rendering must complete in milliseconds on consumer hardware.
In AI generation, the quality of rendering: the detail, coherence, and visual fidelity of the generated output: is determined by the model's architecture, training, and the quality of the inference process. AI rendering artefacts ( blurring, distortion, inconsistency, implausible physics ) are the AI equivalent of traditional rendering artefacts like shadow acne, fireflies, or polygon clipping. Improving AI rendering quality has been a central focus of generative model development, with advances in model architecture, training data, and inference techniques progressively improving the physical plausibility, detail fidelity, and stylistic coherence of generated outputs. For creators, understanding rendering as a distinct phase: the conversion from model-internal representation to pixel image: helps contextualise why generation quality, step count, resolution, and model choice all affect the character of the final rendered output.
