AI image text legibility test: Compare every required character at delivery size

An AI image text legibility test should compare the delivered image with exact approved copy. A word that looks plausible at a glance can still contain an incorrect character, missing punctuation, or an unreadable small line. This protocol defines a human review process; it does not claim a model’s text accuracy.

Freeze the target copy

Save the exact string, capitalization, punctuation, line-break requirements, and any characters that are easy to confuse. Include the intended use, such as a card viewed on a phone or a large presentation slide. Legibility is a delivery requirement, so viewing only a full-resolution source is insufficient.

Separate required text from optional decorative lettering. If a background sign contains arbitrary generated words, decide whether that is acceptable before reviewing outputs. Do not focus only on the headline while overlooking misleading text elsewhere in the frame.

For a fictional brief, the required copy might be OPEN HOUSE followed by a date and a short address. Treat each line as its own review item. These are sample instructions, not evidence that any model rendered them correctly.

Perform two reading passes

First, inspect the asset at its intended display size without zooming. Ask a reviewer to transcribe what they can read. Record missing or uncertain characters rather than letting the target copy supply them from memory.

Second, compare the transcription and visible image against the approved string character by character. Zoom in to locate the error, but do not let successful zoomed inspection override a failure at the actual display size.

Check reading order, line grouping, cropping, contrast, and interference from texture. Correct characters arranged in an ambiguous sequence can still fail the intended message. A price or date deserves particular care because a tiny visual change can alter its meaning.

Use OCR as assistance

An OCR result can highlight candidate errors or help compare long passages, but it is not final confirmation. OCR can miss stylized letters, infer a word from context, or report text that humans cannot comfortably read at delivery size.

Store the OCR output separately from the human judgment if you use it. Mark disagreements and resolve them by inspecting the actual asset. Do not convert a confident OCR result into a claim of guaranteed readability.

The current gateway accepts a prompt string and image-generation parameters; it does not expose a built-in text legibility score. Your review belongs in the application record linked to the archived output.

Release decision

If exact copy is mandatory, consider generating a suitable background and adding approved text through a deterministic layout step. Recheck the final composite after resizing and export.

Accept the asset only when all required strings, reading order, and delivery-size checks pass. Continue with logo placeholders, human rubrics, and negative-space prompts when separating visual generation from exact typography.

Continue with quality and testing

• generation model regression checklist: Maintain a release scorecard with real evidence
• generation prompt revision diff: Review the change as a field-level hypothesis
• AI image color drift test: Control viewing conditions before judging color
• Browse all quality and testing guides