PDF Ai Decompile

Decompile PDF papers back into clean, structured source.

PDF Ai Decompile is a small, cross-platform desktop tool that takes PDF papers and turns them back into formats that are easy for AI tools (and humans) to work with. It is organised around projects (a tabbed workflow you can save and resume) and two activity categories:

• Modify PDF — remove images from a PDF while keeping all text and the exact layout (optionally remove vector figures too, for a text-only PDF), and
• Decompile to Text — rebuild a PDF into LaTeX (one compilable IEEE .tex per PDF plus a shared Latex_Resource folder) and/or Markdown (full text, no images).

Both categories can run together on the files you select. A Passwords tab unlocks protected PDFs (per-file or a shared pool) before processing, and an Inspector tab shows file info, permissions and a page preview. The LaTeX and Markdown outputs are designed so that any AI tool can read the full paper without processing the PDF, and the cleaned PDFs upload without hitting image limits.

Everything you set up — the file list and selection, options, output locations and passwords — is saved into a single .paidproj project file so you can pick up where you left off (Project ▸ New / Open / Save / Save As / Open Recent).

Built with CustomTkinter. Authors: Jerry James & Nisha · Org: Open-Tools-Development · License: GPL-3.0.

Repository: https://github.com/Open-Tools-Development/PDF-Ai-Decompile

---

Project layout

PDF-Ai-Decompile/
├─ README.md                  This file (one level above Scripts)
├─ Doc/
│   └─ SKILL.md               Full architecture / skill document (read this
│                             when migrating to Claude Code)
├─ Published_Tool/            The finished EXE is placed here by build_exe.bat
│   └─ .gitkeep
└─ Scripts/                   All source code and build scripts
    ├─ app/                   UI layer (CustomTkinter)
    │   ├─ __init__.py
    │   ├─ pdf_ai_decompile.py   Main application window + batch runner
    │   └─ about_info.py         Identity, features, how-to, revision history
    ├─ backend/               Backend logic (PDF parsing & conversion)
    │   ├─ __init__.py
    │   ├─ appconfig.py          Per-user config + recent projects
    │   ├─ project.py            Project file (.paidproj) save/load/schema
    │   ├─ pdf_info.py           Scan / password / page-render (Inspector)
    │   ├─ runner.py             Headless project runner (passwords → jobs)
    │   ├─ pdf_common.py         Shared parser (structure, escaping, images)
    │   ├─ pdf_remove.py         Image-removal engine (UI: "Modify PDF")
    │   ├─ pdf_to_latex.py       PDF → LaTeX renderer (4 equation modes)
    │   ├─ pdf_to_markdown.py    PDF → Markdown renderer
    │   ├─ pdf_math.py           Inline-math reconstruction
    │   └─ pdf_equations.py      Display-equation detection + image extraction
    ├─ models/                Native AI models (reserved for future use)
    │   ├─ __init__.py
    │   └─ README.md
    ├─ assets/                Icon + splash and their generator
    │   ├─ icon.ico, icon_preview.png, splash.png
    │   └─ make_assets.py
    ├─ run_app.py             Top-level launcher (also the PyInstaller entry)
    ├─ build_info.py          Build date (auto-generated; reset by clean.bat)
    ├─ requirements.txt
    ├─ install_dependencies.bat
    ├─ run.bat
    ├─ build_exe.bat
    ├─ clean.bat
    ├─ LICENSE
    └─ .gitignore

Quick start (run from source)

1. Install Python 3.8+ (tick Add Python to PATH during setup on Windows).
2. From the Scripts folder, run install_dependencies.bat once.
3. Run run.bat to open the tool.

On macOS/Linux, from the Scripts folder:

pip install -r requirements.txt
python3 run_app.py

Using the tool

1. Project — start a New project (or Open a recent one) and give it a name in the header. Save / Save As writes a .paidproj file holding all of the below so you can resume later.
2. Files tab — Add PDF File(s)… or Add Folder… (optionally Subfolders), then tick which files to process (Select all / Deselect all, or filter by name / path / size / pages).
3. Passwords tab (only if some PDFs are protected) — add a shared password pool and/or a per-file password; Detect passwords now checks them. Locked files are skipped and flagged.
4. Modify PDF tab — enable it, choose Execute or Validate, what to remove, and the output location. When writing beside each PDF a filename suffix (default _noimg) is required so the original is never overwritten; to a separate folder it is optional.
5. Decompile to Text tab — enable it, pick LaTeX and/or Markdown, the equation mode, and the output location.
6. Inspector tab — pick a file to see its info, permissions and a page preview.
7. Click Run. Progress and a log appear at the bottom.

Advanced Modify (Modify PDF tab) also offers: remove restrictions & password, search & replace text (literal or regex), search & replace image (match an embedded image to a reference by similarity, then delete or replace), AI image analysis, a page range to change and which pages to keep, and a Validate mode that reports changes without writing.

Password recovery (Passwords tab) can, for files you are authorised to open, brute force (charset/length/mask, threads, attempt/time limits, files in parallel) and use candidate-generator models — two ship built-in (Markov and a rule mangler, no download) and you can add your own. Confirmed passwords are kept in an encrypted reuse pool. AI models are optional and downloaded on demand into the project folder; without them a heuristic image description and the built-in password models are used.

For LaTeX output, upload the .tex and its Latex_Resource folder to Overleaf, or compile locally with pdflatex (two passes).

Equation handling (PDF → LaTeX)

PDF text extraction cannot fully recover complex LaTeX math. Four modes are selectable in the UI (default: Rebuild as LaTeX math text):

Mode	What it does	Trade-off
Rebuild as LaTeX math text	Editable LaTeX with recovered sub/superscripts and symbols.	Compiles & editable, but complex math is approximate.
Improve inline math only	Recovers inline symbols/subscripts; display equations stay plain text.	Lightest touch.
Hybrid (text + equation images)	Inline math as text, exact image per display equation.	Editable prose + correct equations (as images).
Equation images (exact)	Every display equation inserted as an exact cropped image.	Looks perfect; equations not editable text.

Image file naming

Extracted images use a short, configurable prefix from the PDF name, a unique number (so several PDFs can share one Latex_Resource folder), and the figure/equation number, e.g. RISAidedM_3_Fig-2.png, RISAidedM_11_Eq-5.png. The prefix length is set in the UI (default 9; 0 = full PDF name).

Build a standalone EXE (Windows)

From the Scripts folder, run build_exe.bat. It refreshes the assets, stamps the build date into build_info.py, bundles the icon/splash, and writes:

..\Published_Tool\PDFAiDecompile.exe

A native splash shows while the EXE unpacks. Copy the EXE to any Windows PC — no Python required.

Clean before committing

Run clean.bat to delete build/, any dist/, *.spec, __pycache__/ and *.pyc, and to reset build_info.py. Source files and the EXE in Published_Tool are left untouched.

What the conversion does and doesn't do

• It recovers the text and structure (title, authors, abstract, index terms, all sections/subsections/sub-subsections, figure & table captions, references with \cite{} and an embedded bibliography, author biographies). It is not a pixel-perfect reproduction of the PDF.
• Equations are approximate in text mode; use Hybrid/Image for exact math.
• In many IEEE papers the numeric table grids are vector graphics (not selectable text), so those values are captured as figure images.

See Doc/SKILL.md for the full architecture, module contracts, data flow and extension points.

License

Free software under the GNU General Public License v3.0. See Scripts/LICENSE. It comes with no warranty.