OCR

OCR systems and comparison of results.

The focus is on public domain texts from archive.org.

So far, I think RolmOCR is the best.

OlmOCR did "oftentimes" whereas RolmOCR did the correct "ofttimes".

Main tests were done July, 2025.

New systems added October, 2025.

OCR Systems (October, 2025)

Churro

• https://github.com/stanford-oval/Churro

"Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition"

This is kind of a pain because they use pixi for install, and docker to run vLLM. That can all be worked around, but it's just tedious work. Meh.

I tried using it on Sir Gawain and the Green Knight which had original spelling and characters using a modern font, but it failed on most of the characters, such as thorns. Some Middle English characters were correct.

• https://archive.org/details/sirgawaingreenkn00unse/page/n37/mode/2up

Deepseek OCR

• https://huggingface.co/deepseek-ai/DeepSeek-OCR
• https://github.com/deepseek-ai/DeepSeek-OCR

git clone https://github.com/deepseek-ai/DeepSeek-OCR
cd DeepSeek-OCR/
pyenv local 3.12
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -r requirements.txt
pip install torch psutil
# A6000
export TORCH_CUDA_ARCH_LIST="8.6"
# H100
export TORCH_CUDA_ARCH_LIST="9.0"
# B200
export TORCH_CUDA_ARCH_LIST="10.0"
MAX_JOBS=8 pip install flash-attn --no-build-isolation --force-reinstall --no-cache-dir --no-deps
# Copy over sample script
cp -p DeepSeek-OCR-master/DeepSeek-OCR-hf/run_dpsk_ocr.py foo1.py
# Edit it with image path
vim foo1.py

Not surprising, but this doesn't work on Middle English characters.

OCR Systems (July, 2025)

docext

• https://github.com/NanoNets/docext

docext Install

git clone https://github.com/NanoNets/docext
cd docext/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .

docext Run

This will run a gradio web interface with user/password admin / admin at URL:

• http://localhost:7860

python -m docext.app.app --model_name hosted_vllm/nanonets/Nanonets-OCR-s --no-share

Need to write CLI script.

doctr

• https://github.com/mindee/doctr

doctr Install

git clone https://github.com/mindee/doctr
cd doctr/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .

doctr Run

Need to write CLI script.

EasyOCR

• https://github.com/JaidedAI/EasyOCR

EasyOCR Install

git clone https://github.com/JaidedAI/EasyOCR
cd EasyOCR/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .

EasyOCR Run

easyocr --lang en --gpu True --file test.png --output_format standard --detail 0 --paragraph True > output.txt

• jp2: Yes
• png: no
• tif: no

GOT-OCR2.0

• https://huggingface.co/stepfun-ai/GOT-OCR2_0
• https://github.com/Ucas-HaoranWei/GOT-OCR2.0

Ok, not good enough formatting to bother doing full tests.

InternVL

• https://github.com/OpenGVLab/InternVL
• https://huggingface.co/OpenGVLab/InternVL3-78B
• https://internvl.readthedocs.io/en/latest/

Kraken

• https://github.com/mittagessen/kraken
• https://kraken.re

Kraken Install

git clone https://github.com/mittagessen/kraken
cd kraken/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .
pip install -e .[pdf]
kraken list
kraken get 10.5281/zenodo.2577812

Kraken Run

kraken --input test.png test.txt --format-type image --device cuda:0 binarize
kraken --input test.png test-bin.png --format-type image --device cuda:0 binarize
kraken --input test-bin.png test-segment.json --format-type image --device cuda:0 binarize
# wtf
kraken --input XXX test.txt --format-type image --device cuda:0 ocr

Llama

Llama models and derivatives. Llama4 maverick has some of the best numbers in published benchmarks.

Marker

• https://github.com/datalab-to/marker

Marker Install

Isn't set up to install from git... No pyproject.toml, no setup.py, etc.

mkdir marker-local
cd marker-local/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install marker-pdf[full]

Marker run

marker_single test.pdf --output_format markdown --output_dir output --format_lines --llm_service marker.services.ollama.OllamaService --ollama_base_url http://localhost:11434 --ollama_model llama3.2-vision

• jp2: Yes
• tif: Yes

Both llama3.2-vision and qwen2.5vl:72b produced identical results.

Ollama Fail

It is returning a (mediocre) result, and giving this failure, even though the Ollama server is fine: [WARNING] marker: Ollama inference failed: 500 Server Error: Internal Server Error for url: http://localhost:11434/api/generate

Ollama syslog gives this error (with typo(?) of Ollama name): msg="mllama does not currently support parallel requests"

To (not) fix this error with the llama3.2-vision model, run Ollama with this environment variable:

OLLAMA_NUM_PARALLEL=1

That prevents Ollama from having an error in the logfile, but Marker still gives this error: [WARNING] marker: Ollama inference failed: 500 Server Error: Internal Server Error for url: http://localhost:11434/api/generate

It also prodcues the error with the qwen2.5vl:72b model.

It does not produce this error if --use_llm CLI option is removed.

MinerU

• https://mineru.net/
• https://github.com/opendatalab/MinerU
• https://mineru.readthedocs.io

MinerU Install

git clone https://github.com/opendatalab/MinerU
cd MinerU/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .[all]

MinerU Run

mineru --path test.png --output results --method auto --lang en --formula false --table false --device cuda:0

• jp2: No
• png: Yes
• tif: No

OCRFlux

• https://github.com/chatdoc-com/OCRFlux

OCRFlux Install

sudo apt-get install poppler-utils poppler-data ttf-mscorefonts-installer msttcorefonts fonts-crosextra-caladea fonts-crosextra-carlito gsfonts lcdf-typetools
git clone https://github.com/chatdoc-com/OCRFlux
cd OCRFlux/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .
pip install -e . --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer/

OCRFlux Run

python -m ocrflux.pipeline testdir --data test.pdf  --model ChatDOC/OCRFlux-3B/

• jp2: Yes
• tif: Yes

OlmOCR

• https://github.com/allenai/olmocr
• https://huggingface.co/allenai/olmOCR-7B-0225-preview

Install OlmOCR

sudo apt-get install poppler-utils ttf-mscorefonts-installer msttcorefonts fonts-crosextra-caladea fonts-crosextra-carlito gsfonts lcdf-typetools
git clone https://github.com/allenai/olmocr
cd olmocr/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .
pip install olmocr[gpu]  --extra-index-url https://download.pytorch.org/whl/cu128
pip install https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.5%2Bcu128torch2.7-cp38-abi3-linux_x86_64.whl

Run OlmOCR

# yes --pdfs is used even if it isn't a PDF.
python -m olmocr.pipeline ./output --markdown --pdfs test.png

• jp2: No
• png: Yes
• tif: No

RolmOCR

• https://huggingface.co/reducto/RolmOCR

Using this OlmOCR fine tune model requires using vllm, or hacking up OlmOCR code.

RolmOCR Install

Needs VLLM install.

This repo has a script to use with RolmOCR.

git clone https://spacecruft.org/books/OCR
cd OCR/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -r scripts/rolmocr/requirements.txt

RolmOCR Run

Need to run VLLM and a custom script.

To run VLLM:

export VLLM_USE_V1=1
vllm serve reducto/RolmOCR

Then run the script:

python scripts/rolmocr/bookocr.py --output results/test.txt test.png

• jp2: Yes
• tif: Yes

OpenOCR

Based on Paddle, etc.

• https://github.com/Topdu/OpenOCR

PaddleOCR

• https://github.com/PaddlePaddle/PaddleOCR

PaddleOCR Install

Fails to run from repo build. Doesn't import paddle, etc.

git clone https://github.com/PaddlePaddle/PaddleOCR
cd PaddleOCR/
git checkout $(git tag -l | sort -V | tail -1)
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install -e .

mkdir paddleocr-local
cd paddleocr-local/
python -m venv venv
source venv/bin/activate
pip install -U setuptools pip wheel
pip install paddlepaddle-gpu==3.1.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
pip install paddleocr

PaddleOCR Run

paddleocr ocr -i test.pdf --save_path results --lang en

• jp2: No
• png: Yes
• tif: No

Qwen

Qwen models and derivatives.

• https://github.com/QwenLM/Qwen2.5-VL
• https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct
• https://github.com/QwenLM/Qwen2.5-VL/blob/main/cookbooks/ocr.ipynb

Qwen Derivatives

Optimized for OCR.

• docext.
• RolmOCR.
• https://huggingface.co/syntheticbot/ocr-qwen

Tesseract

• https://github.com/tesseract-ocr/tesseract

Tesseract Install

sudo apt install automake build-essential pkg-config libpango1.0-dev libcairo2-dev libicu-dev libleptonica-dev
git clone https://github.com/tesseract-ocr/tesseract
cd tesseract/
git checkout $(git tag -l | sort -V | grep -v -e rc -e alpha -e dev -e b | tail -1)
./autogen.sh
./configure
make -j$(nproc)
sudo make install
sudo ldconfig
make training
sudo make training-install
cd ../
# 3 gig repo
git clone https://github.com/tesseract-ocr/tessdata_best
sudo cp -p tessdata_best/eng.traineddata /usr/local/share/tessdata/
tesseract --list-langs

Tesseract Run

tesseract test.png results -l eng --psm auto --oem lstm_only

• jp2: Yes
• tif: Yes

Samples

Quixote

Shelton Harvard Classic Editions

Thomas Shelton translation from The Harvard classics Volume 14.

• https://archive.org/details/harvardclassics14eliouoft
• https://archive.org/details/harvardclassicsf0014char
• https://archive.org/details/harvardclassicsv0014edit_f6g1
• https://archive.org/details/harvardclassicsv037564mbp

License

Apache 2.0 License. Unofficial project, not related to upstream projects.

Upstream sources under their respective copyrights.

Copyright © 2025 Jeff Moe.