It's easier than ever to de-censor videos

Line-scan imaging and everyday analogues

• Several comments connect the demo to line-scan cameras and slit-scan techniques used in industrial vision systems and sport photo finishes.
• People note you can approximate the “traveling slit” reconstruction with your own eyes by moving past gaps (e.g., bathroom stall doors), sparking a tangent about US vs European stall design and privacy.
• Rolling shutters and old film shutters are cited as related “moving slit” exposure mechanisms.

Blur, pixelation, and information leakage

• Multiple commenters stress: blur and naive pixelation rarely remove information; they mostly redistribute it. Deconvolution or search over candidate texts can often recover content, especially with known fonts and UI.
• Blur is closer to an invertible convolution; pixelation is likened to a weak hash that can be brute-forced in small regions.
• Larger block sizes and more noise make attacks harder but not always impossible, especially with priors (likely filenames, words, etc.).

Practical redaction techniques and common failures

• Strong consensus: if you really need to hide something, you must destroy the original information, not just cover it visually.
• Recommended patterns: solid opaque shapes, then re-screenshot or print-scan; or rasterize PDFs and verify no text remains.
• Many historical failures are mentioned:
  • Image formats keeping old data (aCropalypse, leftover buffers).
  • Embedded thumbnails or previews not updated after cropping or censoring.
  • PDFs where text is only “black-highlighted” but still selectable.
  • Font metrics and character positioning leaking names even under black boxes.
• Some suggest replacing real content with fake/lorem ipsum, then applying blur/pixelation for aesthetics.

Video-specific issues and mitigations

• The key vulnerability in the article’s example is movement of text under a fixed pixelation grid: multiple frames act like many measurements of the same underlying signal.
• Suggested mitigations:
  • Pixelate once, then overlay a static censored screenshot on all frames.
  • Use pure-color masks or fake-looking but uncorrelated pixelation.
  • Add deliberate noise or scramble patterns, though practicality is debated.

Ethical and legal concerns

• AI “decensoring” of Japanese porn is discussed: some see it as merely generative porn, others call it “deeply unethical” when it violates performers’ expectations or local law context.
• Broader concern: advances in de-anonymization threaten blurred faces/voices in older investigative journalism; French public TV reportedly moved to actors and back-shot filming and has pulled some archival material.

Historical and technical context

• Commenters argue that multi-frame deblurring, blind deconvolution, superresolution, and similar techniques have existed for decades (e.g., astronomy, biomedical imaging); what’s new is accessibility and tooling, not the core math.