PNG to Text: Keep the Background

One lettering mask, three PNG backgrounds

On October 9, 2026, we made three synthetic English images and read them in the live browser tool. All measure 1000 by 420 pixels. We drew the same five lines in Arial at 36 pixels, at the same positions, into one lettering mask. We then used that mask for each export. No image was resized, cropped, or converted into JPG. These are test files, not personal documents.

• png-transparent.png: 24,591 bytes; white lettering over transparency, saved as RGBA. The alpha channel carries the lettering mask; the underlying RGB channels are white throughout.
• png-dark.png: 28,993 bytes; white lettering on an opaque black background, saved as RGB.
• png-light.png: 31,586 bytes; black lettering on an opaque white background, saved as RGB.

These byte counts are measured fixture sizes, not upload limits. The transparent file may look blank against a white viewer background. That appearance is not evidence that its lettering was removed: download it and view it over a dark background to inspect the white letters.

The five source lines were:

PNG background check.
Reference: PNG-204.
Total: 48.50 USD.
Keep the visible background.
Review before copying.

We queued transparent, dark, and light in that order and ran Extract text once. All three entries reached Done. The transparent file's recognized body was exactly:

FING BEckgroune check:
Reierence: PING 204:

ete 48.50 USE,

Keep the visible backgrounol
Review belore copying:

Both opaque files produced this body:

PNG background check.
Reference: PNG-204.

Total: 48.50 USD.

Keep the visible background.
Review before copying.

The transparent result changed words, the reference, punctuation, and the currency label. The number 48.50 survived, which would not make the surrounding field safe to reuse. Both opaque bodies matched the source wording and punctuation, with extra blank lines. The exact combined status was Finished: 413 characters extracted from 3 images. Metadata showed 413 characters · 55 words · 26 lines, including filename headings and spacing. The browser error log was empty.

This is evidence about these files in this run, not proof that all transparent PNGs fail. We changed transparency, background, and color representation together. We did not instrument the engine's decoding stages, so this test cannot establish precisely where the wrong letters arose. It also does not establish that white-on-black is always better than black-on-white: both opaque versions worked here.

Reproduce the check without replacing your work

1. Keep a separate test tab. Save any corrected text from your real task first. Download the three fixtures, open a fresh tool tab, and add them in the listed order.
2. Inspect each filename section. Run Extract text and compare PNG-204, USD, and the final two sentences against the source. A Done label confirms processing, not correct transcription. Do not score the result from its combined character count.
3. Compare a prepared copy of your own PNG. Keep the original unchanged. In a local image editor, place its visible lettering over an opaque contrasting background and export a separate PNG. Preserve dimensions and crop so you are not also testing a resize or a different text region.
4. Choose by the recognized body. Test the original and prepared copy under different filenames. Correct important fields by looking at their source image, then use Copy all or Download .txt. If the prepared copy still fails, stop treating transparency as the only possible explanation.

There is no background, inversion, or flattening button in the current ClearPixel tool. The preparation step happens outside it; this article is not advertising a new image-editing feature. Rerunning extraction replaces the editable result, so keep corrections before another comparison. See the save-before-rerun workflow for that behavior.