Ibteda Digital Library Preserves 1,800 Rare Books, but Ten Human Corrections Beat Bigger AI

Ibteda Digital Library reached Google News after three Pakistani friends put about 902,000 clicks on two budget Nikon cameras while preserving rare Urdu books. Their decade-long effort produced 526,000 photographs of open-book spreads, according to a project account and subsequent reporting.

The striking camera counts made the story travel. Yet the more important result came after daily digitization work stopped. One team member converted years of manual Photoshop edits into training labels for a neural network.

That experiment delivered an awkward finding for an industry conditioned to expect gains from larger models and more data. Extra training books, higher-resolution inputs, and a ResNet-50 backbone did not solve the central cropping problem. Ten corrections from a human operator did.

The project therefore represents more than an unusually determined scanning operation. It shows why cultural preservation cannot simply import the assumptions behind large commercial AI systems. An archive must preserve unusual marks, marginal notes, punctuation, and material defects that a general model might classify as noise.

It also provides a pointed contrast with industrial book-scanning programs. Recent investigations describe companies buying and destructively scanning printed books for AI training. Ibteda worked slowly, preserved physical volumes, and made at least part of the resulting collection available to readers.

The Numbers Behind the Google News Headline

The camera counts tell a story about endurance, but the unfinished processing queue explains why Ibteda had to change course.

Ibteda began around 2015, when its founders were researching historical editions of Ghalib’s Diwan. They struggled to locate important volumes that were out of print, inaccessible in university collections, or held privately.

The team responded by building its own archive in Pakistan. It had no institutional digitization laboratory, commercial scanning bureau, or dedicated equipment budget. Its rig began with a flat table, inexpensive Philips LED lights, a camera mounted above the book, and glass recovered from a photocopier.

A single shutter release captured both facing pages. That method accelerated photography, but every image also contained the tabletop, page edges, gutter shadows, uneven margins, stains, and handwritten additions.

The Nikon D5300 eventually accumulated roughly 576,000 shutter actuations. The D3300 added about 326,000, bringing the commonly reported total to approximately 902,000. Those figures came from shutter data associated with the cameras, not an estimate based only on the number of published pages.

The team photographed about 526,000 open-book spreads. A spread contains two facing pages, so the number cannot be treated as 526,000 completed books or even 526,000 finished page files. Photography was only the first stage.

The project’s more detailed account says operators finished 575,729 individual page crops across 1,765 books. That completed work became the basis for the later machine-learning experiment.

Some coverage rounds the collection to 1,800 books, while current library descriptions refer to a much larger catalog assembled through the broader initiative. These measurements describe different things: processed volumes, photographed material, catalog holdings, and partner-hosted selections.

For example, the Rekhta collection says Ibteda began in 2016 and describes a catalog exceeding 5,000 books and 3 million pages. Rekhta currently presents a selected portion of that wider collection rather than every raw or processed photograph.

That distinction matters because the viral headline can imply a completed conversion of every photographed spread. In reality, the processing burden outlasted the team’s ability to sustain daily operations.

Ibteda reportedly wound down routine work in April 2026 after almost a decade. The group had successfully preserved a substantial body of Urdu literature, but hundreds of thousands of captured spreads still demanded careful processing.

The team’s cameras had survived far beyond the workload expected from an improvised community operation. Human attention became the scarcer resource.

Capturing a Book Was the Easy Part

Ibteda’s bottleneck was not taking photographs. It was turning irregular historical pages into files that an archive could trust.

After each exposure, an operator opened the photograph in Photoshop. The operator separated the facing pages, straightened them, selected appropriate boundaries, and handled stains or other visual interference.

This was not ordinary batch cropping. Many volumes used Nastaliq, a flowing writing style commonly associated with Urdu and Persian literary traditions. Its dots, diagonal forms, compact marks, and diacritics complicate automated cleanup.

A small dark speck might be dirt. It might also be a meaningful dot that changes a letter. A line near the edge might be an accidental shadow, a printed border, or a handwritten annotation worth preserving.

Lithographs added further variation. Pages could contain irregular impressions, aging paper, warped geometry, deep gutter shadows, marginal writing, tables, or decorative elements. No single rectangle represented the correct crop for every volume.

The operator also made aesthetic and editorial choices. One book might look best with a narrow margin. Another needed more surrounding space to retain notes, binding evidence, or an unusual printed border.

Those choices created a hidden variable. The preferred margin was not always visible in the pixels. Two competent archivists could look at the same page and select slightly different acceptable boundaries.

That problem separates archival processing from many commercial imaging tasks. A shopping application can tolerate minor visual normalization. A preservation system cannot casually remove marks simply because they resemble defects.

Ibteda’s archive also had to support future inspection. The workflow reportedly stored completed material in a ZFS pool, a storage system that uses checksumming and related integrity features. BLAKE3 manifests recorded cryptographic fingerprints that could expose later file changes.

This emphasis on traceability is important. A clean-looking page is not automatically a trustworthy archival object. Researchers must know that software did not silently rewrite the source.

The same principle applies to modern knowledge management. Search and automation create value only when people can trace information back to stable, understandable source material.

For Ibteda, every hour saved by automation carried a corresponding risk. A model that erased one stain correctly but removed one Urdu mark elsewhere could introduce a permanent textual error.

The project therefore needed more than an image-enhancement model. It needed a conservative system that knew when to propose a change, when to preserve the original pixels, and when to send a page back to a person.

That requirement drove the team toward separate models and processing rules for cropping and retouching. It also explains why a large volume of historical Photoshop work became valuable training data.

A Decade of Photoshop Edits Became Training Data

The team’s most reusable asset was not its camera hardware. It was the record of 575,729 human decisions made during ordinary production work.

Each finished Photoshop page encoded a choice about where the page began and ended. However, those decisions did not initially exist as structured labels ready for machine learning.

The team had to connect finished pages back to their original photographs. Its published technical discussion describes using SIFT and MAGSAC during that recovery process.

SIFT, or Scale-Invariant Feature Transform, identifies distinctive visual features that can still match after changes in scale or rotation. MAGSAC is a robust estimation method that helps reject incorrect matches while fitting image geometry.

Together, these techniques allowed the project to estimate how a completed page related to the raw camera image. Conservative acceptance rules filtered questionable matches rather than forcing every file into the dataset.

The resulting geometry became supervision for a crop-prediction model. In practical terms, the team converted years of human editing into examples showing where operators had placed page boundaries.

This is a valuable pattern for other small archives. An organization may lack a formally labeled AI dataset while possessing years of edits, corrections, approvals, and production outputs. Those records can contain more useful domain knowledge than a generic image collection.

However, historical workflow data is not automatically reliable. It can include inconsistent operator preferences, changes in standards, duplicated files, failed edits, and undocumented exceptions.

Ibteda’s case was especially challenging because some inconsistency represented legitimate judgment. A narrow crop could be correct for one volume, while the same margin could damage another.

The crop model still learned visible structure. It could recognize probable page boundaries, gutters, and surrounding table space. Yet the team found that its remaining errors were often nearly constant within each book.

That pattern was revealing. The model was not simply failing to detect pages. It was missing the operator’s preferred inset for a particular volume.

In a technical discussion, the project author said expanding training from 378 to 572 books did not improve performance on unseen volumes. A ResNet-50 model fitted the training data better but stayed flat on held-out material.

Increasing the input resolution to 1,024 pixels also failed to unlock the expected gain. A spatial prediction head did not resolve the problem either.

These negative results matter because they challenge a familiar AI development reflex. When performance stalls, teams often add data, compute, resolution, or a larger architecture.

That strategy works when the missing signal exists somewhere in the training inputs. Ibteda’s experiments suggest the decisive preference was not visible in a new book’s pixels.

No larger backbone can reliably infer information that the image does not contain. The system required a small amount of new human context.

Ten Corrections Beat a Bigger Neural Network

Ibteda improved its crop model by calibrating it to each book, not by asking a larger network to guess an invisible preference.

For a new volume, an operator corrected ten predicted crops. The system compared those corrections with its original predictions and calculated the median residual, meaning the typical offset between machine output and human preference.

It then applied that adjustment to the remaining pages in the book. Pages within one volume usually shared paper dimensions, printing style, binding geometry, and an operator’s desired margin.

The project reports that this calibration increased its pass@80 measure from 0.71 to 0.83 on held-out volumes. Pass@80 represented the share of books for which at least 80 percent of pages met the project’s acceptance criteria.

That gain came from ten operator corrections, not a larger model. It supports a practical form of human-in-the-loop automation, where a person supplies limited examples that guide the system through a specific batch.

The mechanism is modest compared with a general-purpose vision model. That is precisely why it fits the task. It targets the stable, book-level bias that the model cannot observe directly.

The system does not eliminate the archivist. It moves the archivist’s attention to the beginning of a volume, then applies that judgment consistently across later pages.

This approach also limits the cost of failure. If calibration looks wrong, the operator can stop processing that volume. A fully autonomous system might repeat the same subtle cropping mistake across hundreds of pages before anyone notices.

The result has implications beyond books. Many document workflows contain preferences tied to a customer, project, collection, instrument, or reporting period. Those preferences are often absent from the raw content.

A general model can identify visible structure, while a handful of corrections provide local policy. That combination can outperform a larger model trained across incompatible preferences.

Ibteda’s retouching work followed an equally conservative philosophy. A U-Net, a neural architecture designed for pixel-level image segmentation, identified candidate regions containing stains, stamps, or unwanted marks.

The model handled detection only. Classical OpenCV routines reconstructed paper texture inside the approved mask, while pixels outside that area remained unchanged.

Training labels used three states: REMOVE, KEEP, and IGNORE. REMOVE identified marks considered safe to erase. KEEP protected content that resembled noise but belonged to the document. IGNORE withheld ambiguous regions from training.

The project author reported that stricter labeling increased mark intersection-over-union from 0.56 to 0.60. Intersection-over-union measures the overlap between a predicted region and its human-labeled target.

More importantly, the stricter process reportedly reduced false removal of Urdu diacritics to zero in the evaluated material. The team treated any erased diacritic as a deployment veto, regardless of the model’s average score.

That rule captures a central preservation principle. A high aggregate metric cannot excuse deletion of meaningful text. One culturally significant failure can matter more than thousands of correctly cleaned pixels.

This Preservation Model Challenges Destructive Scanning

Ibteda used AI to extend an archive’s life, while industrial scanning programs often optimize for acquiring text as quickly as possible.

The contrast became sharper during August 2026. Investigations reported that technology companies were buying physical books, removing their bindings, scanning the pages, and discarding the remains.

Court records previously showed that Anthropic bought and destructively scanned millions of books while assembling an internal research library. Anthropic has said its acquisition programs did not target rare or antiquarian books.

More recent reporting focused on Amazon. A tracked shipment reportedly reached an Amazon facility where workers cut book spines before scanning. Amazon did not provide public answers about the program’s full scale or how scanned text would be used.

The industrial scanning investigation raised concerns among booksellers because some purchased titles appeared scarce. The evidence does not establish that every bulk buyer intentionally targets unique or irreplaceable copies.

Still, the underlying incentive is clear. Large AI developers want substantial quantities of long-form human writing. Removing a book’s binding allows individual pages to move rapidly through an automated scanner.

Ibteda faced the opposite objective. Its physical books were cultural objects, not temporary containers around training text. The project needed to preserve layout, handwriting, printing artifacts, and other evidence beyond the words.

That distinction affects every engineering decision. Destructive scanning prioritizes speed, flat pages, and clean text extraction. Archival imaging must balance readability with material fidelity.

Ibteda also distributed access rather than locking the scans inside a proprietary training pipeline. A selection is visible through Rekhta, while related material has been preserved through the Internet Archive collection.

Open access does not resolve every copyright or stewardship question. Digitizing out-of-print material can involve complicated rights, privacy, and ownership issues. Community archives still need policies governing access, takedowns, donor expectations, and sensitive records.

Nor does Ibteda’s workflow prove that every fragile book can safely tolerate a glass platen. Conservation requirements vary by binding, paper condition, ink, and historical value. Some volumes require professional cradles, lower pressure, or specialized imaging.

The system’s machine-learning findings also remain project-reported results. The code and weights were still under archival review when the author discussed them publicly. Independent teams have not yet reproduced the entire pipeline across different scripts and collections.

The evaluation metrics reflect Ibteda’s own acceptance criteria. Another archive might choose different margins, error thresholds, or definitions of an unacceptable textual change.

These limitations do not erase the contribution. They define what must happen before others treat the workflow as a transferable preservation standard.

The strongest claim is narrower than the Google News headline. Ibteda found a promising way to recover supervision from historical edits and combine general predictions with per-book human calibration.

What Archivists and AI Teams Should Watch Next

The project’s value now depends on reproducibility, cross-collection testing, and evidence that its safeguards survive routine use.

The first signal is a public technical release. The team has described training recipes, label-recovery thresholds, routing rules, and a reproduction contract. Releasing code, weights, evaluation samples, or a carefully governed dataset would let others inspect those details.

A useful release must include difficult examples, not only successful pages. Researchers need gutter shadows, handwritten notes, dense Nastaliq, damaged lithographs, stamps, borders, and pages where automation was rejected.

Independent reproduction would strengthen the project’s central finding. If other teams also discover that a few local corrections beat indiscriminate model scaling, the result could influence many specialized document systems.

Failure to reproduce it would still be informative. The improvement might depend on Ibteda’s capture rig, operator consistency, book formats, or particular acceptance metric.

The second signal is testing across institutions and scripts. Urdu preservation is the project’s core use case, but related challenges appear in Persian, Arabic, Ottoman Turkish, and South Asian manuscript collections.

Cross-collection trials should measure more than crop overlap. They should examine lost diacritics, altered marginalia, incorrect page order, accidental text removal, operator time, rejection rates, and the cost of reviewing failures.

The most valuable benchmark would separate visible geometry from editorial preference. That design could test whether calibration works because it captures a true volume-level pattern rather than compensating for one dataset’s quirks.

The third signal is operational adoption. A promising model does not automatically clear a backlog. Archives need interfaces for entering corrections, reviewing uncertain pages, tracing transformations, and restoring originals.

Operators should be able to see exactly what the model changed. They also need immutable source images and recorded parameters for every derived page.

Ibteda’s separation of neural detection from constrained reconstruction offers a useful starting point. It limits where changes can occur and creates a clearer audit boundary than unrestricted image generation.

Future experiments may test diffusion-based restoration, where a generative model fills damaged regions. That approach could produce visually convincing paper, but archival teams would need guarantees against changes outside an approved mask.

A realistic standard should therefore reward abstention. Systems must be allowed to say that a page is ambiguous and requires a person. High automation rates are less valuable when they conceal irreversible textual errors.

The project also offers a broader lesson for AI product teams. Domain expertise often appears as corrections, exceptions, and preferences accumulated during real work. Treating those traces as structured feedback can be more effective than continually enlarging a model.

For archivists, the next step is not replacing judgment. It is identifying where ten careful decisions can safely guide the next thousand pages.

For readers who discovered Ibteda through Google News, the camera counts are an impressive entry point. The enduring question is whether institutions will test, fund, and reproduce the workflow before similar collections disappear into inaccessible shelves or private training datasets.

Seek out the available Ibteda collection, compare a processed page with its material complexity, and watch for a public technical release. If another archive can reproduce the 0.71-to-0.83 calibration gain while protecting every meaningful mark, this small project will have delivered something larger AI systems often miss: automation that adapts to human judgment without erasing the evidence that judgment was meant to preserve.