Keep the Font and Background: What Image Text Editing Can Really Preserve
Actual serif and patterned-background tests explain approximate font matching, pixel restoration, 2K scaling and the limits of changed-area checks.
“Keep the same font” sounds like a font lookup. In this editor it is a different process: the image model sees the lettering in a picture and attempts to redraw the replacement in a similar visual style. It does not identify, install or license the original typeface.
This distinction matters when a result is used for a logo, printed packaging or strict brand work. An attractive approximation may still have the wrong serif, counter shape, baseline or spacing. If exact typography matters, use the actual licensed font and the source layout.
A controlled serif replacement
Our source contained BOLD ROAST, set as large display lettering on an owned flat graphic. We requested DARK ROAST. The returned phrase is legible and the broad serif appearance is similar in this particular example.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 35.2 seconds observed, 3.9% pixel-difference estimate. The serif appearance is similar; no font file was identified and exact glyph or kerning identity is not established.
The difference estimate is 3.9%. That number describes pixels that exceeded the browser check’s threshold, not how closely a typeface was matched. The result has no font-identification score, and the test does not establish that the new characters share identical outlines with the original font file.
When reviewing a replacement, look at repeated letters first. Compare stroke weight, curved counters, the tops of capitals and the space between letter pairs. Then inspect characters that were not present in the old phrase: those require the model to infer shapes from limited evidence.
What background restoration actually does
The model is asked to change only the named phrase. That request can fail: a generative model may redraw the surface, alter colors or shift the framing. The browser therefore compares the result with your uploaded original.
When the image shape is close and the changed area is contained, the compositor keeps the original outside a softened mask around the differences. The soft boundary helps avoid clipping new strokes. This is a pixel-based heuristic, not an understanding of which pixels belong to a word.
A small color shift can fall below the difference threshold. A compressed or rescaled image can introduce differences that are unrelated to text. When the measured change is broad, the editor warns that it cannot safely restore the original automatically.
A repeating pattern is a harder case
We also tested WEEKEND SALE to SPRING SALE over diagonal lines. The target phrase is present, but the local pattern is not an exact reconstruction. This is a case to examine closely rather than present as a flawless edit.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 40.7 seconds observed, 12% pixel-difference estimate. The requested phrase is legible, but the diagonal pattern close to the new letters is not an exact reconstruction.
The difference estimate increased to 12%. Even though the heuristic still categorized the change as contained, manual review found local texture differences. That is why a “restored” status must not be interpreted as a guarantee of visual acceptability.
Repeating patterns, knitwear, brickwork and edges passing behind text are difficult because the original hidden surface is unknown. The model supplies an estimate. If exact continuity is required, source artwork or a careful manual reconstruction is more reliable.
Resolution and preservation are separate questions
The Pro model is requested at its 2K tier. Most landscape outputs in this small set were 2528 × 1686 pixels from 1200 × 800 inputs. The editor retains that larger working size for download rather than shrinking it back to the source dimensions.
The unedited background is nevertheless derived from the original input when restoration is applied. Scaling a small source does not add new captured detail. At the edge of an edited word, generated and scaled source pixels can have different sharpness or grain.
Check both the typography and the background at full size. Then reduce the image to its intended display size. A seam that is visible at 400% may be irrelevant for a tiny web graphic, while a wrong letter remains a wrong letter at any size.
A practical acceptance checklist
- Read the complete replacement, not just its general shape.
- Inspect other text, logos and straight edges near the edited region.
- Compare repeated letters and their spacing.
- Turn on the change highlight and inspect the mask boundary.
- Check actual output dimensions and any low-resolution warning.
- Keep the untouched original and reject the result if the change is too broad.
Our same-font page and background-preservation page include the relevant tests and capability boundaries. The performance page explains why technical completion, timing and visual quality are reported separately.