JPG to Text: Test the Copy You Actually Have

Same canvas, different JPEG exports

On October 8, 2026, we tested two generated JPGs in the live tool. Each starts from the same white RGB canvas, measuring 1000 by 420 pixels, with black Arial text drawn at 36 pixels. The text positions, background, and dimensions did not change. We saved each file directly from that canvas with Pillow, using JPEG quality settings of 95 and 5 and the same subsampling setting. Neither file was made by resizing or recompressing the other.

jpg-check-q95.jpg is 47,303 bytes. jpg-check-q5.jpg is 11,465 bytes. These are measured file sizes, not upload limits. The quality settings are encoder inputs, not percentages of OCR accuracy. The downloadable pair contains no personal document data.

Both source images contain exactly these five lines:

JPG extraction check.
Order reference: ABC-123.
Total: 48.50 USD.
Keep the original image.
Review before copying.

We added the quality-95 file first and the quality-5 file second, then ran Extract text once. Both queue entries reached Done. Beneath each filename heading, the result was:

JPG extraction check.
Order reference: ABC-123.
Total: 48.50 USD.

Keep the original image.
Review before copying.

The code, decimal point, wording, and punctuation matched in both sections. The engine inserted the same blank line after the total in each. The exact status was Finished: 281 characters extracted from 2 images. The metadata was 281 characters · 38 words · 15 lines. Those combined counts include both filename headings and spacing; they are not an accuracy score. The browser error log was empty during this run.

The smaller export did not produce a wording error in this particular test. We therefore cannot present these fixtures as proof that compression always breaks OCR, or that a large JPG always works better. Large, upright black letters on white are a narrow case. A phone photo, colored label, small receipt, or repeatedly shared image is a different input. This experiment changes encoding settings while holding the underlying canvas fixed; it does not test every factor involved in a real photo.

Repeat the pair, then inspect your own JPG

1. Use a fresh tool tab. Download both fixtures and add them in filename order. Keep any real task in a separate tab so this check cannot replace text you are still editing.
2. Run the same queue. Select Extract text and locate each filename section. Check ABC-123 and 48.50 USD separately, then the surrounding sentences. Do not treat the combined character count as evidence that those fields are correct.
3. Compare the body, not the headings. Different filenames make the overall sections different even when the recognized words match. Blank lines also affect the displayed count. Compare the actual source sentences rather than expecting two identical metadata totals.
4. Test the file you intend to use. Add your own English JPG in another fresh tab and check its important fields against that exact image. A passing synthetic control shows the tool can read that control, not that your document has been validated.

If you have an original and a shared copy, keep their names distinct and compare both. Record whether dimensions, crop, and visible lettering also changed. Otherwise an observed difference cannot be attributed to JPEG compression alone. Save corrected text before another extraction run: this tool replaces the result when it reruns the queue. For that behavior, see the save-before-rerun guide.