A generated image fails in three places: the pixels, the file and the metadata.

Generated images arrive oversized, because the model renders at a resolution nobody resized before publishing. A 2048-pixel square used as a 300-pixel thumbnail costs weight and delays first paint.

They also arrive with unpredictable contrast. Diffusion output is often mid-grey dominant, with a narrow range between the darkest and lightest areas, which looks flat on a bright screen and worse once a CMS recompresses it.

The third failure is metadata. Provenance markers, colour profiles and XMP blocks are stripped by resizing tools, by social platforms and by the copy-paste that turns an asset into a screenshot of an asset.

The console decodes the file in the browser and reports what it can read from the bytes and the bitmap.

A checker can report what markers a file carries, which is useful in both directions. A manifest present means someone took care of provenance. Absent on an image the team knows was generated means something removed it.

Markers are text inside a binary, so absence is weak evidence. The checker states what it found and does not claim a file without markers is unedited, or a file with them is authentic.

Treat the result as an inventory: which provenance exists, which was lost, and whether the loss happened at export, at resize or at upload. That last question is answerable by checking the file at each step.

An image used as content needs a description that conveys what a reader would take from it. An image used as decoration should carry an empty alt attribute, not a sentence.

Generated images tend to be illustrative rather than evidential, and alt text should reflect that: describe what is depicted, avoid inventing claims the picture does not support.

Placement is a quality question too. A hero image whose subject sits behind a headline overlay is a rendering bug waiting for the smallest viewport.

No measurement of weight, contrast or metadata can say whether the image is appropriate, on-brand, accurate or legal to use. A technically clean file can still depict something a compliance team would refuse.

Style consistency across a set is a judgement, and so is whether an image that looks photographic is being used in a way that implies it is documentary.

For anything published as evidence or under a client's name, the verdict belongs to a person who looked at the image and at the context.

Check the exported file before it reaches the CMS, not after. Resize to the display size, then re-export, then check again, because the resize is where metadata disappears.

Record the outcome for anything used in a regulated or client-facing context: dimensions, weight, markers present, and who approved the use.

Then keep the source file. Every downstream recompression is another chance to lose the manifest, and the only way to prove what the original carried is to still have it.