Create an annotated before-and-after comparison
"Looks better" is a weak review. A useful comparison points to what changed, says whether anyone asked for it, and gives the next instruction that would fix it.
What you'll make
a one-page comparison: the original and the edited image side by side, numbered circles on each change, and a short note for each saying keep, remove or fix.
What you'll need
the original image, the AI-edited version, the instruction you gave for the edit, a slide or design tool (PowerPoint, Google Slides or Canva) and your work AI account. About 15 minutes.
Example result.
| Where | What changed | Next |
|---|---|---|
| 1 Sky | Grey to clear blue (asked for) | Keep |
| 2 Stall sign | “Quayside Market” now reads “Quayside Marke” (not asked for) | Fix |
| 3 Upper left | A second flag has appeared (not asked for) | Remove |
1. Put the two images side by side
Make one slide or page with the original on the left and the edit on the right, at the same size. Size matters: a bigger image looks more detailed simply because it is bigger. Write the edit instruction above them, so everyone knows what was asked for.
Why this works: with the instruction on the page, every difference is either something you asked for or something you didn’t.
What you should see: both images at the same size, the instruction above them, and room for notes underneath.
2. Ask AI to list every visible difference
Upload both images to your work AI account and ask for a list, not a verdict. Then look yourself, one area at a time: corners, text, faces, logos and edges are where unrequested changes hide.
The first image is the original. The second is an AI edit made with this instruction: [instruction].
List every visible difference between them, including ones the instruction didn't ask for.
For each, give its location, what changed, and whether the instruction asked for it.
If an area is too small or blurred to compare, say so instead of guessing.[instruction] is the exact edit request you gave.
Why this works: “every visible difference, including ones the instruction didn’t ask for” stops the AI checking only the part you mentioned.
What you should see: a numbered list of locations and changes, each marked as asked for or not.
3. Number the changes and write the next request
Add a numbered circle on the edited image for each difference you confirmed. Under the images, write one line per number: where, what changed, asked for or not, and keep, remove or fix. Then turn the remove and fix notes into one narrow request.
Using the original as the reference, fix only these: [numbered notes].
Keep [what to keep]. Don't change anything else, including text, faces and logos.[numbered notes] are your remove and fix lines; [what to keep] is the change you asked for and liked.
Why this works: numbered notes make the request specific, and “don’t change anything else” protects what’s already right.
What you should see: a new edit to compare with the original again, using the same numbers.
Check the result
- Every numbered circle has a matching note, and every note names a place in the image.
- Each note says whether the change was asked for.
- Text in the image, such as signs, labels or names, has been read letter by letter.
- After the fix, the whole image was compared again, not just the numbered areas.
The usual failure: the review covers only the change you asked for. The notes say “sky improved, looks great”, and nobody spots that the sign now has a missing letter. Spot it when every note on the list is something you requested.
Your list only covers the change I asked for. Compare the two images again, area by area:
the four corners, any text, any people, logos and the edges of the main subject.
List every other difference you can see, with its location.Worked example: a market banner with a new sky
Fictional example. A town events team asked AI to brighten the sky in a banner photo for the Quayside Market.
Instruction: Replace the grey sky with a clear blue one. Change nothing else.
Original: grey sky, one flag on the left pole, stall sign "Quayside Market"
Edit: blue sky, two flags on the left pole, stall sign "Quayside Marke"First review (wrong): “1. Sky: now clear blue, as requested. The image looks brighter and more inviting.”
Repair: the repair prompt above. The second pass finds the extra flag in the upper left and the missing “t” on the sign.
Repaired notes: the three-row table at the top of this page, then the step 3 request: “Fix only these: 2, restore the sign text ‘Quayside Market’; 3, remove the second flag. Keep the blue sky.”
- The sign was the costly miss. A banner that misspells the event’s name gets printed at two metres wide.
- “Change nothing else” in the original instruction didn’t stop the drift. Checking did.
- After the fix, compare the whole image again. Fixing the sign can shift the stall’s edges or colours.
Try it
Zoom in and note what changed
Ask AI for evidence-tied annotations rather than a generic judgement.
A useful note names a place in the image, says whether the change was asked for, and ends in keep, remove or fix.
The example in text
- 1 Contact shadow: looks natural, keep
- 2 Plant upper right: not asked for, remove
- 3 Rim and handle: reshaped, check against the original
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Make the notes readable without the circles
Write each note so it makes sense on its own: “2, stall sign, lower centre: ‘Market’ missing its final ‘t’”. Someone reading on a phone, or with a screen reader, should get the conclusion without matching circles to numbers. Colour and arrows can support a note, not replace it.
Keep the review separate from sign-off
An accurate comparison doesn’t approve the image for use. Brand, rights and permission decisions belong with whoever owns them, especially for photos of real people. See team portrait approval.
Use the same method for charts and documents
Number each change and tie it to a source value or rule rather than appearance. A chart whose axis label lost its unit is the chart version of the missing “t”. The chart label guide applies this to units and context.