AI Image Artefact Taxonomy

Definition. A classification of the ways AI image processing fails visually, such as halos, seams, over-smoothing, fabricated text, repetition, colour shift and geometry distortion, used to diagnose and to build targeted checks. Without a taxonomy, quality problems are described vaguely and cannot be fixed.

How it works. Define categories from observed defects. For each, define a detectable signal where possible: halos show as high-frequency ringing near edges, seams as a discontinuity at tile boundaries, fabricated text as a mismatch between detected and original text. Build checks for the detectable ones and rely on sampled review for the rest.

Example. A pipeline's output is described as looking off. Applying a taxonomy shows 41 percent of complaints are over-smoothing, 28 percent fabricated text and 19 percent halos. Over-smoothing is traced to an aggressive denoiser, halos to sharpening before resizing, and fabricated text to the upscaler. Each is fixed with a targeted change, and the complaint rate falls 76 percent.

Use case. An artefact taxonomy is valuable for any image pipeline, and it converts vague quality complaints into actionable defects. It is also the basis for automated quality checks.

Warnings. Vague quality complaints cannot be fixed, so always classify. Never assume one cause for a mixed set of complaints. And never rely on metrics alone for defects that are visually obvious but metrically invisible.

Limitations. Some artefacts are hard to detect automatically, requiring sampled human review. The taxonomy also needs revision as models introduce new failure modes.

← Browse all glossary terms