Visual difference triage
Compare screenshots between builds with a model of what humans perceive, so only meaningful visual changes are flagged — and group them so one change in a shared component is reviewed once.
How it works
Plain pixel comparison of screenshots is noisy: anti-aliasing, font rendering, a one-pixel shift and dynamic content all register as changes unless thresholds and masks are tuned carefully. Perceptual comparison, which has been in commercial testing tools for years, ignores differences a person wouldn't notice and highlights the ones they would. That part is mature. Newer models add a second layer, and it is this layer that is still emerging: describing what changed ("button moved below the form", "price now wraps to two lines"), recognising regions that are expected to vary such as dates and ads, and grouping differences across many screens that come from the same underlying change. Reviewers approve or reject a group rather than each screenshot, and approved changes become the new baseline.
When to use it
Products with many screens, themes, breakpoints or browsers where visual regressions are common; design-system changes that ripple across the product; and teams whose screenshot tests were abandoned because of noise.
Limitations
The tool decides what is significant, and its threshold will sometimes hide a change that matters — a slightly wrong colour on a warning, a misaligned decimal. Approving grouped changes in bulk makes it easy to wave through a real regression with a batch of intended ones. Visual comparison checks appearance only; a screen can look identical and behave differently.
Watch for
The model marks the price area as dynamic content, so a price that later renders as zero is masked on every run. Review the regions it proposes to ignore as you would code, and keep prices, warnings and legal text out of them.
What you still own
Approving baselines, and deciding which visual changes were intended.
Cite this
Qlarify Labs. (2026). Visual difference triage. Retrieved from https://labs.qlarify.fi/ai-in-testing/visual-difference-triage