What to do when translated image text is missing, distorted, or unreadable

When translated image text is missing, distorted, or unreadable, stop the publishing workflow. Identify the symptom, compare every text region with the source, and decide whether a changed input supports another attempt. If the meaning cannot be verified, reject the result or redraw the affected area manually. Do not assume that another run will fix it.

Treat troubleshooting as a diagnosis with a stop condition. The useful outcome is not always a repaired image. Sometimes the correct outcome is a documented rejection.

Identify the failure type

Name the problem before changing anything. "Bad text" is too vague to guide a reviewer or designer. Separate the visible symptom from the possible cause.

A result can have more than one symptom. Record them separately. Missing text is a coverage failure. Wrong wording is a language failure. Broken letters are a rendering failure. Crowding is a layout failure. This classification prevents one correction from hiding another problem.

Do not infer the intended word from shape alone. If a fluent reviewer has to guess, the text is unreadable and the asset fails.

Check whether source quality caused it

Inspect the source before blaming the output. Small type, low contrast, compression, overlapping effects, a busy background, or text cropped near an edge can make the source harder to interpret. This does not prove why a failure occurred, but it tells you whether a cleaner input is available.

Check the source at full size and at the intended crop. Confirm that every word is visible, the text is not already distorted, and no important character blends into the background. Make sure you are using the approved source rather than a compressed preview or an old export.

If an editable source exists, consider exporting a cleaner version with legible text and enough space for the target copy. If only a flattened image exists, document that limitation. Do not pretend the image contains font files, layers, or editable coordinates when it does not.

Source cleanup supports a new attempt only when it addresses a visible input problem. It is not a guarantee. A clear source can still produce an unusable result, and a reviewer must check the whole image again.

Compare every original text area

Build a text-region table with one row for each source phrase instead of scanning the image casually.

Use the table to catch omissions, duplicates, substitutions, and text attached to the wrong object. Read the output in its natural order, then compare it in the region order. Both views matter. The region check finds missing content; the reading-order check finds confusing hierarchy.

Inspect names, numbers, and claims separately. A generated image may look plausible while one digit, unit, or proper name is wrong. Those details require an approved source and the right owner, not a visual guess.

Save annotated crops for every failed region. A precise note such as "bottom right label, second word unreadable" is actionable. "Text looks weird" is not.

Retry only when appropriate

Another attempt is reasonable when you can make a controlled change and state what it is meant to address. Examples include using a cleaner source export, providing shorter approved target copy, increasing space around the text, or choosing a workflow with manual layout control.

Before retrying, preserve the failed output and review notes. Change one major input at a time when possible so the next reviewer knows what differed. After a new result arrives, repeat the full language, visual, and factual checks. Do not review only the region that failed before because other details may have changed.

Do not retry merely because the first result failed. Stop when:

• the wording remains uncertain;
• the same critical region keeps failing;
• the subject or product details are being damaged;
• approved copy cannot fit without losing readability;
• the asset contains high-risk facts that cannot be verified;
• the team lacks permission or an authoritative source.

There is no universal retry count. The stop condition depends on risk, review authority, and whether a different input or workflow can address the problem.

Reject or redesign when meaning is unsafe

Reject the output when a reasonable reviewer cannot verify the message. That includes missing warnings, changed claims, unreadable names, uncertain numbers, and text that appears connected to the wrong object. Visual polish does not make those failures acceptable.

Manual redesign may be safer when editable source files exist, the copy is high-risk, the layout has little room, or repeated reconstruction changes image details. A designer can place approved text deliberately and keep a record of the source. That workflow still requires language and factual review.

Text removal may be an option when the image does not need embedded copy and the surrounding page can carry the message. It is not suitable when the text is required for identification, instruction, or another purpose the owner has approved.

Document the final decision with the failed regions, reviewer, source of truth, chosen next action, and any unresolved risk. If the asset is rejected, keep the record. A clear rejection prevents the same unsafe output from returning later without context.