# Visual regression testing without the noise

> Screenshot comparison finds the breaks nobody wrote an assertion for, and drowns them in font rendering and animation. How to keep the signal.

Published 2026-10-08, about FlowQA.

Visual regression testing works when three things hold: baselines come from the same rendering environment as the runs, visual drift is reported separately from functional failure, and a person approves a new baseline from a run they have looked at.
Without those three, the comparison reports every font hint and every half-finished animation, and the team soon stops reading it.

## What visual checks catch that assertions do not

A functional test asserts what it was told to.
It does not notice that the sidebar now overlaps the content, that the dark theme lost its contrast, or that a button is rendered off-screen but still "visible" to the test.
A screenshot of each step, compared against a known-good one, sees all of that.
That is the signal.

## Where the noise comes from

**Rendering differences.**
The same page on a Mac and on a Linux worker renders fonts differently.
Line breaks move, buttons shift a pixel, the diff lights up.
Comparing a baseline from one environment against a run from another produces noise on every run.

**Motion.**
A screenshot taken mid-transition differs from one taken at rest.
Spinners, fades, carousels.

**Dynamic content.**
Dates, counters, avatars, anything that changes between runs.

**Everything at once.**
A full-page diff of a dashboard with ten widgets will differ somewhere on most runs.

## Keeping the signal

**Compare like with like.**
A baseline is valid for one rendering profile: operating system, browser build, viewport, and the fonts installed.
When the worker image changes, the profile changes, and the baselines are re-approved, deliberately.
Never compare a cloud run against a screenshot from a laptop.

**Settle before shooting.**
Wait for the network to go quiet and animations to finish.
A tool that captures on a fixed delay will catch motion sooner or later.

**Report drift as drift.**
A run where every assertion passed and three screenshots differ is a different thing from a run where a step failed.
Show it as such.
A failing test is a stop; visual drift is a review.
Mixing them in one red mark is what makes people ignore both.

**Approve from a run you looked at.**
A new baseline should come from a completed run on the current version of the tests, after a person has seen the differences and agreed they are intended.
A run with functional failures cannot be a baseline; a run whose only differences are drift can, once reviewed.

**Mask what moves.**
Dates and counters are masked or asserted separately.

## A worked example

*A composed example: the team, the timings and the counts show the shape of the work and are not measurements from one customer.*

A team adds visual checks to the twelve journeys in their merge gate.
The first week is loud: every run reports drift on every page, because the baselines were captured on a developer's Mac and the runs are on Linux workers.
They re-approve the baselines from a cloud run, and the noise stops.

Two weeks later a run passes every assertion and reports drift on the settings page.
The diff shows the save button has moved under the footer at the tablet viewport after a CSS change.
No assertion would have caught it; the button was still clickable to the test.
They fix the CSS, the next run shows no drift, and nobody had to approve anything.
