IMAGE & PHOTO ANALYSIS

1. Overview: Images in the Epstein Files

The official Epstein file releases contain approximately 900,956 files across 12 datasets. Nearly every one of these files is a PDF — and inside those PDFs are embedded images: photographs, screenshots, scanned pages, charts, and hundreds of thousands of blank white rectangles.

Our pipeline has extracted and individually classified 949,123 images from these PDFs, sorted them into 47 content categories using an AI vision model, and initially flagged 546,976 blank-like pages by pixel analysis. After OCR/reclassification of blank-labeled buckets, the current confirmed blank bucket is 17,831 pages (~1.9% of all extracted images).

2. Why Are Images Stored as PDFs?

This is one of the most important questions about the release format, and the answer says more about the releasing agency than it does about the files themselves.

What happened

Photographs, screenshots, scanned documents, text messages, business cards, and handwritten notes were all individually wrapped in PDF containers before release. A simple 200KB JPEG photograph becomes a 1.5MB single-page PDF. A screenshot of a text conversation that could be a PNG file becomes a PDF that requires a PDF reader to open.

What the original files were

These images originally existed as standard image files — .jpg, .png, .tiff, .bmp — or as embedded images within email clients, litigation databases, and forensic imaging tools. They were converted to individual PDF pages as part of the Concordance/IPRO litigation support workflow used by the DOJ and cooperating law firms.

Why this matters

• You can’t search images inside PDFs. Image-based PDFs don’t have searchable text. Every page requires OCR (Optical Character Recognition) before it becomes searchable — a process that takes seconds per page and hours at scale.
• You can’t thumbnail or preview PDFs easily. Standard file browsers can display image thumbnails instantly. PDFs require rendering, which makes visual scanning of large collections impractical without specialized tools.
• File sizes balloon. The PDF wrapper adds overhead to every file. Across 900,000+ images, this adds up to hundreds of gigabytes of wasted space.
• Metadata is lost. Original image files contain EXIF data — camera model, date taken, GPS coordinates, orientation. PDF conversion strips all of this. For evidentiary photographs, this metadata loss is significant.
• It fragments multi-page documents. A 47-page scanned deposition becomes 47 individual PDF files, each containing a single page image. Reassembling the original document requires knowing which page numbers belong together — information that is often only available in the proprietary Concordance load file.

3. The Extraction Pipeline

To make the images accessible, we built a multi-stage pipeline that extracts, classifies, and organizes every embedded image from every PDF in the release.

Filename Convention

Each extracted image gets a structured filename that encodes its classification and source:

deposition-transcript+adult-woman+affidavit__EFTA01609671_20260130_p001_i001.png

• Primary category + secondary tags — what the AI thinks the image contains
• EFTA ID — the original Bates/document identifier from the release
• Date, page, image number — extraction metadata

Technical Details

4. AI Classification (CLIP)

Every extracted image is classified using CLIP (Contrastive Language–Image Pre-training), an AI model developed by OpenAI that can match images to text descriptions without being specifically trained on any particular category. We use the ViT-B/32 variant.

How it works

CLIP compares each image against a set of 48 custom text labels and returns a probability score for each one. The labels were specifically designed for this type of legal/evidentiary content:

Accuracy & Limitations

• CLIP is good at broad categorization — it reliably distinguishes email screenshots from deposition pages from photographs
• It struggles with similar document types — an affidavit vs. a legal filing vs. a court document can look nearly identical, so these categories have overlap
• Multi-label tagging helps — each image gets its top 3–5 labels encoded in the filename, so an image classified as email-screenshot+contact-list captures both aspects
• Human review is still needed — the AI classification is a sorting tool, not a final determination. Critical images should always be verified

5. Content Breakdown: What’s in the Images

Based on CLIP classification of 949,123 extracted images, here is the high-level breakdown of what the Epstein files actually contain visually:

Content Type Distribution

What this tells us

6. Full Category Inventory

All 26 active categories with current file counts. Categories were created by the CLIP classification pipeline and have been consolidated through manual review.

Complete Table

7. Blank Pages & Waste

Blank handling now has two distinct numbers: the initial image-level flags from pixel sweep, and the current confirmed blank bucket after OCR/reclassification. The initial sweep flagged 546,976 pages as blank-like; after reprocessing blank-labeled folders, the remaining confirmed blank bucket is 17,831 pages.

The category table below is a historical image-level sweep snapshot and should be read as pre-reclassification detail, not the current blank residual.

How blank detection works

Each image is loaded as a pixel array. If the mean pixel value exceeds 252 (out of 255) and the standard deviation is below 3, the image is classified as blank. This catches pure white pages, near-white pages, and pages with only faint scanner artifacts. The threshold is conservative — genuinely blank pages typically have mean > 254 and std < 1.

Blank Page Breakdown

8. Cleanup: Duplicates, Tiny Files, Corrupt Images

Beyond blank pages, the extraction process identifies several other categories of problematic files that inflate the release without adding content.

What each category means

Total waste calculation

Adding up all identified waste: 17,831 blanks + 250,243 duplicates + 3,554 corrupt + 697 tiny = 272,325 junk files out of 949,123 extracted images. That’s 28.7% of extracted images. After removing this identified waste, approximately 676,798 pages remain.

9. Which Datasets Contain What

The 12 datasets vary enormously in their image content. Based on our scanning and classification work so far, here is what we know about the visual content of each dataset:

Where the images are concentrated

Three datasets (DS9, DS10, DS11) account for 98.4% of all files. The remaining nine datasets combined are a rounding error in terms of image volume.

10. Current Processing Status

All major pipeline stages are complete. Here is the final state of each:

📸 Interactive Image Galleries

Browse classified images directly in your browser — filterable, searchable, with full-size lightbox viewing.

THE PDF FILES

Image analysis based on extraction and CLIP classification of 949,123 images from 900,956 PDFs.
Classification: OpenAI CLIP ViT-B/32. Extraction: PyMuPDF. Blank detection: Pillow + NumPy.
All scripts and methodology are open source and reproducible.

← Back to Main Site  ·  File Analysis & Findings  ·  Gallery Picks  ·  Handwritten Notes  ·  Cast Hub  ·  Search Files