How to stop a generated graphic inventing the numbers and labels on it

Last updated 2026-09-15 · 8 min read

What is the short version?

A generated diagram wants to look complete. Given a refinery it will add pipe labels; given a chart it will add an axis with values; given a document prop it will write a name on it. None of those came from you.

1. Tell Vizard Agent the graphic must contain no text.
2. Supply the figures and labels yourself.
3. Ask for every card proofread at full size.

What do you need before you start?

Your actual numbers and the labels you want, written out. Vizard Agent can only refuse to invent a figure if there is a real one to use instead, so a request for a stat card with no stat in it will get one filled in.

• The real figures. Exact, as they should appear.
• The labels. Every one that should be visible.
• What must have no text. Diagrams, props, backgrounds.
• Where the numbers came from. Useful for checking.
• The typeface. So the overlay matches the rest.

What do you type into Vizard Agent?

Say "no text in the image" to Vizard Agent as a separate instruction of its own. Sitting next to a description of the graphic you want it reads as redundant, and it is nevertheless the single sentence that changes what comes back.

Prompt

Variants worth knowing:

• "No text, numbers or labels at all." The instruction that works.
• "Do not add any others." Closes the gap you left.
• "Generate it blank." For props and documents.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on a real explainer whose stat cards came back carrying figures nobody had supplied. Rebuilding the cards is the easy half of the job; it is the proofreading pass afterwards that stops the next one going out with a mistake in it.

1. Verifies the key numbers against the source before anything is built.
2. Reviews the generated graphics as a set.
3. Rebuilds the diagrams with no invented labels in them.
4. Proofreads the text in every graphic.
5. Proofreads the stat cards and callouts separately.
6. Regenerates the cards that failed, text-free.
7. Checks one card frame by frame for a clean version.
8. Fixes a duplicated entry on that card.
9. Inspects the document prop and the end card at full size.
10. Generates the document still blank, with no names on it.

Step one is worth doing even when you are confident. A figure you half-remember and a figure the generator invented look identical on a card, and the only difference is whether you can point at where it came from.

Step eight is the failure nobody expects. Generated cards repeat entries — the same line twice with slightly different wording — and at playback size it reads as a list rather than as a mistake.

Step ten is the version of this that catches people out most. A prop document is background dressing, so nobody proofreads it, and then someone pauses the video and reads a name that does not exist.

What does the result look like?

Cards that say only what you said. Every figure on screen is a figure you supplied, the labels are your labels, the props are blank, and Vizard Agent tells you which graphics it had to rebuild and exactly what it found when it proofread the finished set at full size.

When does this not work well?

Some graphics are mostly text, and a text-free version of them is nothing. A chart with no axis labels, a comparison table with no headings — for those the whole graphic has to be built as typography rather than generated at all.

• Graphics that are mostly text. Build them as typography.
• Handwriting or signage in a scene. Hard to keep blank.
• Very small text. Illegible either way; better left out.
• Charts with real data. Build from the data, not from a picture.
• Non-Latin scripts. Generators mangle them; overlay instead.

How do you fix a result that came back wrong?

Quote the text you can see. Vizard Agent can find it in the frame and either remove it or replace it with the correct value, and a quoted string is unambiguous in a way that "the number is wrong" is not.

• "There is a figure I never gave you." That card is regenerated text-free.
• "The same line appears twice." The duplicate is located and removed.
• "The document has a name on it." Regenerated blank.
• "The label is misspelled." Replaced with typography you supplied.

How does Vizard Agent compare to doing it yourself?

By hand you prompt for a stat card, get something that looks professional, and use it. The invented figure is not obviously invented — that is the whole problem — so it survives every review until somebody who knows the real number sees it.

Common questions

Why does it invent figures at all? Because a card with an empty space looks unfinished. Vizard Agent removes the space instead.

Can I just tell it the right number? Yes, and also say no others. Otherwise Vizard Agent may add more alongside it.

What about charts? Build those from the data. Vizard Agent will not draw a chart as a picture.

Will the typography match? Yes. Vizard Agent sets your text in your typeface over the artwork.

What about a document in the background? Ask Vizard Agent to generate it blank. Someone always pauses.

Can it proofread what I already have? Yes. Send the cards and Vizard Agent inspects each at full size.

What is a duplicated entry? The same line twice, worded slightly differently. Easy to miss, easy to find.

Does this apply to maps? Yes, and especially — place names are invented freely.

Can it check my figures too? Yes, if you say where they came from. Vizard Agent verifies against the source.

What about logos in a generated scene? Same rule. Ask Vizard Agent for the scene without branding.

Will removing text make it look empty? Sometimes, which is why the typography goes on afterwards.

What about other languages? Overlay rather than generate. Vizard Agent handles the script properly that way.

Does this cost more? A little, because the failed cards are regenerated. Far less than a correction.

How do I know it proofread them? Ask Vizard Agent what it found. A clean pass is reported as a clean pass.