Starstruk Logo MarkStarstruk

v1.2.0

datalab-to/markerv1.2.0Jan 2, 2025by VikParuchuri

AI Summary

This release introduces an optional LLM mode to boost output quality for tables, math, and complex pages. It also improves OCR heuristics for automatic error detection and re-OCRing, while significantly speeding up the layout model.

Key Highlights

  • Optional `--use_llm` flag improves tables, inline math, forms, and complex pages.
  • Automatic detection of bad OCR text with re-OCR capabilities based on new heuristics.
  • Layout model is now ~2x faster and more accurate.
  • New CLI flags for stripping existing OCR and disabling image extraction.

New Features

  • LLM mode (`--use_llm`) for improved output quality.
  • Automatic bad OCR detection and re-OCRing.
  • LLM-based image captioning.
  • New OCR heuristics model.
  • Easier extraction of individual block types from documents.

Full Release Notes

# Overview

Significant improvements to quality and speed.  There is now LLM mode, which will optionally leverage LLMs to boost output quality.  OCR heuristics are significantly improved, and marker will now make good decisions about when to re-OCR the document.  Layout model is faster and more accurate.

### Quality

- Optionally pass the `--use_llm` flag to improve tables, inline math, forms, complex pages, and general quality.
- Automatically detect bad OCR text and re-OCR the document.  This consists of some PDF-level heuristics and a new OCR quality model.
- Pass the `--strip_existing_ocr` flag to always ignore existing OCR and redo it instead.
- Layout blocks are now detected more accurately when passing `--use_llm`.

### Speed

- Layout model is now half the size and ~2x faster (most of the runtime in the general case is layout, so this should result in a big overall speedup).  It's also more accurate.

### Misc

- Pass the `--disable_image_extraction` flag to avoid extracting images.  
- Pass `--use_llm` and `--disable_image_extraction` to automatically convert images to descriptions.
- Made it easy to extract individual block types from the document (for example, getting all tables out)

## Partial Changelog
* Add New OCR Heuristics Model by @tarun-menta in https://github.com/VikParuchuri/marker/pull/427
* Vik dev by @VikParuchuri in https://github.com/VikParuchuri/marker/pull/434
* High Quality Layout Builder and Text Processors by @iammosespaulr in https://github.com/VikParuchuri/marker/pull/429
* Vik dev by @VikParuchuri in https://github.com/VikParuchuri/marker/pull/438
* Vik dev by @VikParuchuri in https://github.com/VikParuchuri/marker/pull/447
* Additional heuristics for bad PDF text extraction by @iammosespaulr in https://github.com/VikParuchuri/marker/pull/446
* LLM based image captioning by @VikParuchuri in https://github.com/VikParuchuri/marker/pull/454

## New Contributors
* @tarun-menta made their first contribution in https://github.com/VikParuchuri/marker/pull/427

**Full Changelog**: https://github.com/VikParuchuri/marker/compare/v1.1.0...v1.2.0