February 15, 2026
PDF goes brrr, DjVu fans go grrr
DjVu and its connection to Deep Learning (2023)
The book format AI legends made—then the web ditched it and fans are furious
TLDR: DjVu, a book-scan format by AI pioneers, often beats PDF for small, clear pages. Commenters clash over why it faded—pirate stigma, weak support, and DRM—while others tout it for text recognition, highlighting how file choices decide what knowledge is easy (or hard) to share.
A blog love letter to DjVu—the scrappy book-scan format crafted by AI heavyweights Yann LeCun, Léon Bottou, and Yoshua Bengio—set off a nostalgia bomb. The post argues PDF botches scanned text (hello fuzzy JPEG), while DjVu keeps pages clear and tiny. Cue the comments: book nerds came out swinging, reminiscing about underground libraries and side-eyeing today’s bloated PDFs. One fan called DjVu “the best thing for scanned books,” while others confessed they always thought it screamed “pirate ware.”
That’s the drama: is DjVu’s downfall a tech fail or a vibe fail? qdotme blames the “sketchy software” aura; stared says they never saw it used legally; qingcharles laments that most DjVu files get converted to PDFs because nobody knows or supports the format; and joecool1029 drops a spicy accusation that a major archive ditched DjVu for DRM (digital locks), not quality. Meanwhile, nico_h pipes up with a practical take: DjVu is perfect for OCR (turning scans into selectable text). The memes wrote themselves: “PDF goes brrr, text goes blur,” and “AI dads invented the best ‘pirate’ e‑book format.” Whether you’re team minimalist scans or team mainstream convenience, this Locklin on science post turned into a full-on custody battle over the internet’s bookshelf.
Key Points
- •DjVu models documents as mixtures of text and images, enabling high compression while preserving readable text and plots.
- •Early PDFs often stored scans as bitmaps or JPEG/TIFF images, which are suboptimal for text due to transform artifacts.
- •DjVu was co-created by Yann LeCun, Léon Bottou, and Yoshua Bengio, with contributors including Patrick Haffner and Bill Riemers.
- •LeCun and Bottou also created the Lush programming language used in their 1990s research; later ML tooling moved toward Torch with Lua and Python.
- •In the late 1990s, slow internet and large compressed PostScript files made sharing papers difficult; DjVu addressed this by enabling efficient online distribution of scanned documents.