How AI Can Unmask Pixelated Text: Inside the Depix Project

After fixing graffiti‑masked images, even thick‑coded text is no longer safe; the Depix GitHub project now exposes deliberately obscured "text passwords" using AI.

Depix quickly gained popularity, reaching over 7 K stars within three days and surpassing 10 K stars shortly after.

It can recover text hidden behind pixelation (mosaic) that many users think is secure, such as simple WeChat screenshot doodles.

By adjusting image editing parameters (exposure, brightness, contrast, etc.) to extreme values, even seemingly protected doodles can be reversed, but AI now makes this reversal trivial.

Depix – Open‑Source Thick‑Code Text Recovery

Depix uses AI algorithms to restore text that has been pixelated, targeting images created with linear box filters.

The tool is developed by information‑security consultant Sipke Mellema and currently supports English letters, digits, and punctuation.

python depix.py -p images/testimages/testimage3_pixels.png -s images/searchimages/debruinseq_notepad_Windows10_closeAndSpaced.png -o output.png

Full workflow:

Crop the pixelated rectangle from a screenshot.

Paste the corresponding De Bruijn sequence into an editor with matching font settings.

Capture a screenshot of the sequence to create the search image.

Run the command above with the pixelated image and the search image.

Algorithm Overview: Small‑Block Matching with De Bruijn Sequence

The algorithm splits the mosaic area into many small blocks and matches each block against a pre‑generated character library based on a De Bruijn sequence.

For each block, the algorithm pixelates all blocks of the search image and checks for direct matches. It assumes a single match is correct, validates surrounding multi‑match blocks geometrically, and repeats the process to refine results.

The project is not intended for malicious data theft; instead, it demonstrates weaknesses in ECB‑style masking and encourages stronger information‑protection techniques.

Related AI Upsampling: Duke University's PULSE

PULSE (Photo‑Upsampling via Latent Space Exploration) can upscale extremely low‑resolution images (e.g., 16×16 pixels) by 64× to produce realistic 1024×1024 outputs, revealing details such as pores, eyelashes, and hair that were previously invisible.

The technique is generic and could improve image quality in fields ranging from medical imaging to astronomy.

Project links:

Depix: https://github.com/beurtschipper/Depix

PULSE: https://github.com/adamian98/pulse