Compare what changed rather than relying on memory
Compare Images places two local still pictures into a dedicated viewing workspace. Use it to inspect a compressed copy, a resized version with matching dimensions, a retouched image or another deliberate edit. Human visual memory is poor at subtle side-by-side differences; having the two actual files available makes the discussion more concrete.
Choose Image A and Image B as supported PNG, JPEG or WebP. Give each role a clear meaning, such as original and delivery copy. The tool does not decide which one is correct or better. A visible change may be intentional, an encoding artifact or a color-handling difference, so interpret it in the context of the operation you performed.
Start with the simplest view
Side by side is useful for overall composition, subject size and obvious changes. It lets you see both images simultaneously without moving a divider. Use it first to confirm that you selected the right files and that their orientation and framing are appropriate for the comparison.
A slider reveals one image on either side of a movable boundary. Move it across a feature you care about, such as hair, a product edge or small lettering. This can make a local difference easier to notice than alternating between distant thumbnails, especially when the overall pictures look almost identical.
Overlay varies how the images combine visually. It can help identify a shifted edge or slightly changed alignment, but an attractive blend is not itself an assessment of quality. Check whether the pictures actually represent the same crop and pixel grid before treating a ghosted boundary as a processing defect.
Numerical difference needs identical dimensions
The pixel-difference mode compares RGBA values on matching grids. The images must have identical dimensions for a meaningful numerical result. If one image was resized or cropped, the tool cannot infer a geometrical registration that makes every pixel correspond to the same scene point.
Use Resize Image deliberately when the task calls for matching grids, but understand what that adds: resizing creates new pixels and becomes part of the comparison. It does not restore the original correspondence of an earlier crop or account for camera movement. A shifted photograph can show a large difference even when both photographs look good independently.
A difference image highlights where decoded values disagree. It does not know whether those differences are acceptable, important or even perceptible in the final use. A tiny numerical change across a large smooth region and a small but important change in a digit have different meanings to a viewer.
Use threshold and amplification to investigate
The difference threshold helps focus attention on changes above the selected value. Raising it can make minor differences less prominent; lowering it can reveal more subtle variation. Do not raise the threshold merely to make an edit appear unchanged. Record the setting when a numerical comparison forms part of a review conversation.
Amplification makes differences easier to see. It changes the visualization, not the underlying edited file. A dramatic amplified image can therefore exaggerate the visual significance of a small encoding change. Return to a normal view before deciding whether the delivery copy is acceptable.
Inspect alpha as well as color when transparency matters. A graphic can have similar visible RGB values against one background while differing in edge alpha. Try the actual output in the intended design too, because a comparison workspace is not every possible compositing environment.
Ask a question that the comparison can answer
For compression, inspect text edges, gradients, hair and detailed textures. You are looking for damage that affects the image's purpose, not insisting that a lossy delivery copy must be numerically identical to its source. A JPEG created with a new quality setting should be evaluated as a tradeoff between bytes and usefulness.
For a privacy edit, look at the marked region and the whole remaining frame. Numerical similarity outside a box does not prove that all private information has been removed. A successful redaction needs an appropriate output method and its own verification, not simply a difference map showing that some pixels changed.
For color grading, inspect difficult highlights, shadows and skin tones. A large difference can be the intended look, while a smaller difference can still be unpleasant. This tool is not a color-calibration instrument, aesthetic ranking service or guarantee of how a different screen will display the images.
Compare decoded images, not file authenticity
The workspace evaluates supported decoded image content. Two files can display identically while carrying different metadata or container bytes. Conversely, two differently encoded images can have small pixel differences while serving the same practical purpose. Use File Inspector & Hash when exact byte equality with a known digest is the question.
Metadata is also a separate layer. Image Inspector reads supported fields and properties, whereas a visual comparison does not authenticate capture information or detect every embedded identifier. Keep these evidence types distinct when explaining what you reviewed.
Color decoding and browser behavior can affect the displayed result. The tool does not establish universal profile or HDR preservation. If a professional receiving workflow has strict color requirements, verify the actual files there rather than treating one browser comparison as a complete interoperability test.
Boundaries and a sensible conclusion
Each supported still image is limited to 20 MiB and the applicable 12-megapixel image budget. Identical dimensions are required for numerical difference. Available memory and decoding support remain relevant, particularly because a comparison needs resources for both images and its views.
Creative files are processed locally and are not overwritten. Application assets still load from the site. Before leaving, keep the actual source and result files you intended to compare; an on-page view is not a saved review record or a backup. Conclude with the specific finding you observed, such as readable text at the intended size or a visible edge shift, rather than a broad claim that one file is universally better.