Editing Copy Inside a Finished Graphic

The offer ends Friday, the flyer says the old price, and the person who built the original file in whatever design tool made it is unreachable, on vacation, or simply moved on eight months ago. This happens constantly — a promotion updates, a date rolls forward, a price changes by five dollars — and the default response for most teams is still the most expensive one available: rebuild the graphic from the source file, if the source file can even be found. It's the single most requested small edit in marketing production, and it's been oddly hard to do quickly, for a specific reason worth understanding before you reach for a fix.

Two different jobs, one name

"AI can read and write text now" gets treated as a single capability. It isn't — it's two, and they get evaluated by completely different benchmarks because they're doing different work.

Text rendering is the generation-time skill: given a prompt, spell the requested word correctly and place it legibly in a scene the model is creating from nothing. This is what improves when a provider reports gains on an internal text benchmark — it's a measure of "does the model get the letters right when it's building the image," full stop.

Text editing is a different job entirely: given a finished image that already exists, change one specific piece of copy — a price, a date, a headline — while leaving the font, size, color, position, and everything else in the graphic untouched. A model can be excellent at the first and mediocre at the second, because rendering fresh text into an empty region and surgically replacing text that's already baked into a flattened image are not the same operation, even though both involve "the model produces correct text."

Mixing the two up is exactly why the default has been "rebuild it." A strong text-rendering model asked to "edit" a finished graphic will often just regenerate the whole scene around the new text, which is indistinguishable from a rebuild — you've traded one slow process for another one wearing an AI label.

What the benchmarks actually measure

Worth being specific about this, because it's easy to read a text-rendering number as a promise about the edit job. Ideogram reports a 0.97 score on the X-Omni English OCR accuracy benchmark, putting it ahead of every other open-weight release on text rendering despite a comparatively compact model size — a genuinely strong result, and one that describes generation quality: text created as part of a new image, spelled and placed correctly. Microsoft's MAI-Image-2.5 reports an overall +75 point improvement over its predecessor, with the single largest category gain in Text Rendering at +107 points — again, a generation-quality claim, measuring how well the model writes text into images it's producing.

Both are real, both are useful to know, and neither one is a claim about opening an existing flattened flyer and swapping one number without disturbing the rest of the layout. ImagineArt 2.0 is a clean example of where this generation-side strength lives in practice — its catalog listing carries Text-rendering as a named feature alongside Photoreal and 2K output, squarely a text-to-image tool for producing a new graphic with accurate copy baked in from the start, not an editor for a file that already exists.

The actual edit job is a scope problem

Once you're past the "which skill do I need" question, changing copy inside a finished graphic is fundamentally a narrow-scope edit: you want the model to touch a small region — the price string, the date — and leave literally everything else, including the parts of the layout immediately around it, alone. That's a preservation instruction as much as a change instruction, and it needs to be written as both. "Change the price to $39" tells the model what to write. It doesn't tell the model to leave the font, the color, the kerning, and the background untouched — and without that second half, an edit-capable model will often take small liberties nobody asked for, because nothing constrained it not to.

The fix is specific, not clever: name the change and the preservation in the same instruction, every time. "Change the price from $49 to $39. Keep the same font, size, color, and position — do not change anything else in the image." That's the difference between an edit and an accidental regeneration wearing the same file extension.

Running the edit in Versely

For this exact job, Microsoft's MAI-Image-2.5 Edit is built for it directly — its own description calls out pixel-level editing that specifically includes text, alongside cleanup and background work, rather than being a general edit model that happens to touch text as a side effect. A concrete pass:

1. Attach the finished graphic as the reference image for an image-to-image edit.
2. Name the exact change and the exact preservation together — the price, the date, or the headline you're updating, plus an explicit instruction to leave font, size, color and position alone.
3. Generate through an edit-tagged model, not a general text-to-image one — the distinction from earlier matters here directly: you want an editor, not a strong-at-rendering generator that will happily rebuild the whole frame around your new text.
4. Compare old and new side by side at full zoom, specifically along the edges of the changed region — a shifted baseline, a slightly different weight, or a color a few points off from the original are the tells that give away an edited price faster than the wrong number would.

For anything with brand type involved rather than a generic system font, cross-check the result against your actual font files — an edit model's idea of "the same font" is a visual approximation, not a guarantee it matched your specific typeface, and that gap shows up first in a price tag or date stamp, exactly the small, high-scrutiny text most likely to prompt this whole edit in the first place.

What actually goes wrong, and what to check for

Symptom	What it means
New text is a slightly different weight or size	The model approximated the font rather than matching it exactly — check against your real brand font file
Background behind the new text looks subtly different	Scope leaked past the text region into the area around it — tighten the preservation instruction or use a masked edit instead
Kerning or baseline shifted	Common when the new string has a different character count than the old one — worth a manual nudge even after a clean model pass
Color is close but not exact	Compare hex values directly rather than trusting a visual match — small drift is common and easy to miss at a glance

None of these mean the edit failed outright. They mean the review step isn't optional — a copy edit that looks right at a glance and is subtly wrong at the pixel level is worse than an edit that's obviously unfinished, because it's the one that ships.

FAQ

Is a model with strong text-rendering scores automatically good at editing existing text?

Not necessarily. Text rendering measures how accurately a model writes text while generating a new image from scratch. Editing existing copy inside a finished graphic is a different operation — it requires preserving everything around the change, which a strong generator isn't specifically tuned to do unless it's tagged as an edit model.

What's the safest way to change a price or date without altering the rest of the graphic?

State the change and the preservation in the same instruction — name the exact new text, and explicitly instruct the model to keep font, size, color and position unchanged. Run it through an edit-capable model rather than a general text-to-image generator, then compare the result against the original at full zoom before shipping it.

Why does an edited price sometimes look almost right but not quite?

Usually a font-matching approximation, not an outright error — the model gets close to your original typeface without necessarily matching it exactly, which shows up as a slightly different weight, spacing, or baseline. This is most visible on short, high-scrutiny strings like prices and dates, which is exactly the text most likely to need this kind of edit.

Can I use the same model to both generate a new graphic and edit an existing one?

Sometimes, but check which mode you're actually invoking. Many providers ship separate generation and edit variants of the same underlying model precisely because the two jobs behave differently — using the edit-tagged version for an existing file, rather than the general text-to-image version, is what keeps the rest of the layout from being silently regenerated.

Is it faster to just rebuild the graphic in a design tool instead?

Only if the source file still exists and whoever's editing has access to the original layers. When the source is missing or the original brand fonts and assets aren't on hand, a scoped pixel-level edit on the finished export is usually faster — and increasingly the only option if the editable file is genuinely gone.

Try it on your next price update: attach the finished graphic to Versely's AI photo editor, name the change and the preservation together, and check the result at full zoom before it goes out.