Make the comparison fair
A product can occupy different amounts of a preview window even when two files have the same dimensions. One preview may include a large transparent margin; another may zoom tightly to the object. If you compare those fitted views, the smaller label will naturally carry fewer displayed pixels. Begin with the actual downloaded files and the same source detail, not two screenshots of differently sized interface panels.
Record image dimensions in pixels, then open a comparable crop of the identifying word at a native view. A native comparison means judging the underlying image samples rather than an arbitrary fit-to-window size. High-density screens and application zoom conventions can complicate the visual meaning of a percentage, so compare within the same capable editor and inspect actual file information as well.
Four stages where softness can appear
| Stage | Question | Next step |
|---|---|---|
| Original photograph | Was the label sharp before removal? | Recapture if the required word is absent |
| Tool preview | Is the product displayed smaller or resampled? | Inspect the export before changing the source |
| Downloaded export | Are dimensions or actual lettering different? | Compare equal crops and trace processing |
| Storefront rendition | Did the store select a smaller file or new crop? | Repair delivery or layout, then recheck |
MDN documents smoothing when images are scaled in a canvas. That establishes why a display preview can differ from an underlying file; it does not prove which renderer caused a particular CartVisual observation. A softer appearance needs a stage-specific comparison before a cause is assigned.
Worked example: a tea-tin variant line
This is an illustrative file-comparison exercise. A seller starts with a 2400 by 2400 photograph containing “EARL GREY 80 g.” After background removal, the fitted preview looks soft. The downloaded PNG is also 2400 by 2400. In the same editor, equal lettering crops show the same character shapes, while the preview had displayed the tin at half the apparent size. The evidence in this scenario points to display scale, not lost source lettering.
A second delivery file has been reduced to 600 by 600 for an email. Its fine label is less readable. That is a different file with less spatial detail. Do not use the email copy as the master for the store's zoom image. Retain the full-size master and derive a destination-specific file. The email can use a clearer product role or nearby accessible wording where appropriate, without pretending the small raster label communicates every line.
A third hypothetical export changes “80 g” to “60 g.” That is a content mismatch, not softness. Reject it and inspect whether the user chose a generative route rather than a local cutout. Increasing sharpness cannot repair the wrong quantity.
Run a source-to-export check
- Keep the original photograph untouched.
- Download the actual result instead of saving a screenshot of the preview.
- Read the two files' dimensions and format.
- Compare the same opaque interior word and a boundary region at equal scale.
- Open the final destination rendition and repeat the required readability check.
Interior wording and boundary pixels can behave differently. A mask may preserve the opaque label while damaging a thin rim. Conversely, a clean silhouette does not prove that an export retained all interior detail. Keep those findings separate. If a useful file comparison shows no change but the storefront looks soft, do not repeatedly remove the background; investigate which file the page actually serves.
Use the appropriate CartVisual route
CartVisual Background remover is the local cutout route. The separate white-background preservation purpose uses the first source image. General Product images generation may redraw small lettering and should be reviewed as a newly generated depiction. The route descriptions were checked in application source for this draft, not through a new upload-and-export quality experiment.
If you need an exact captured label, choose the source-preserving path that fits the task and compare its result. Do not interpret the word “preserved” as a guarantee about scaling, colors, boundaries or final delivery. Record which operation produced each file so a later reviewer does not compare a generated candidate with a cutout as though they were the same process.
Define a release condition that matches the image role
An identification image passes when the selected product and required wording are accurate at the stated use. A detail image intended to communicate the quantity needs that line readable in the actual delivered rendition. A small campaign thumbnail may communicate the product more broadly while the page provides exact information separately, but it must not display a contradictory visible quantity or flavor.
Keep a comparison record with source file, operation, export file, pixel dimensions, observed character changes, intended placement and unresolved questions. “Preview looks blurry” is a starting observation. “Master label unchanged at equal scale; email rendition too small for the quantity role” is an actionable result. Repair only the failed stage, then inspect the final composite once more.
If the original label is out of focus, no display test can recover the missing evidence. Obtain a clearer source. If the export's native detail changed without an understood reason, keep publication on hold and use the original in a controlled workflow. The purpose of the diagnosis is to avoid needless regeneration while still rejecting genuine product changes. Ask a second reviewer to read the exact variant word without seeing the source transcription first. Then compare that reading with the approved artwork. This is a useful manual cross-check, not an automatic optical-character-recognition result or an audience comprehension study.
Continue with the next decision
Sources and scope
The worked decisions are illustrative editorial exercises, not generated-image tests or measured campaign results. The practical checks in this guide are the seller’s responsibility.
- MDN: scaled canvas image smoothing: Scaled display can use smoothing; a preview is not a native-pixel comparison.
- MDN: image formats: PNG supports transparency; JPEG uses lossy photographic encoding without an alpha channel.