Drop images here. Any text they contain is bent, striped, cut, recoloured and noised until machine reading fails, while a person can still make it out. Every result is checked against Tesseract, PaddleOCR and the strength is raised until none of the original words come back. Only the areas around the text are touched. All metadata (EXIF, GPS, camera info) is removed.
Drop images here or click to choose
JPG, PNG, WebP, GIF, BMP, TIFF, HEIC · up to 40 MB each
Processed files
Nothing processed yet. Files you process appear here and can be downloaded by clicking their name. They are deleted automatically after 24 hours.
How it works
- Find the text. Tesseract 5 scans the image twice (page layout + sparse text) to locate every line and the words it can read confidently.
- Break it for machines, not people. Around each line: hue and saturation drift, random stroke break-up, a hue-shifted ghost copy, wavy strike-through lines in the text colour, hairline cuts in the background colour, a fine mesh at maximum, a sinusoidal warp, chroma noise and speckles.
- Verify. Three gates, cheapest first: Tesseract 5 (four passes: normal, sparse, denoised + upscaled, median-filtered), then PaddleOCR PP-OCRv5 server models (a deep detector + recogniser of the same class as the big cloud services), then Amazon Textract on the actual output file when it is configured. In Auto mode a level is retried with a new random seed, then raised, until every engine reads nothing.
"OCR-proof" means every listed engine read none of the original words on the final file. Google Vision is not called. Higher levels are harder for people too: Auto stops at the first level that passes, so re-run at Maximum or Extreme if you want extra margin. Images with no detected text get a light, nearly invisible treatment and are marked as such.