💻EasyOCR

Not every OCR job needs a vision LLM in the loop. Sometimes you just want text pulled off an image, locally, without an API key, a cloud bill, or a network request. This node wraps EasyOCR - the well-known open-source OCR library from JaidedAI - directly, and it's the honest "it just works offline" option in comfyui_LLM_party, compared to the pack's cloud-vision-backed Mini Advanced OCR or the heavier local GOT-OCR2 model.

How it works

EasyOCR detects text regions in the image and recognizes the characters in each region, using detection and recognition models specific to the language(s) you pick. It runs entirely on your machine - GPU if you have one and want the speed, CPU if you don't.

The inputs and outputs that matter

• image - the image to run OCR on.
• gpu (default true) - run detection/recognition on GPU. Needs a working CUDA-enabled PyTorch install; flip it off if you're on CPU-only hardware.
• language_list - pick from EasyOCR's supported languages (44 options: English, Simplified and Traditional Chinese, Arabic, and dozens more). Only pick the languages actually present in your image - every extra language adds recognition overhead and, on first use, another model download.
• is_enable - the pack's standard bypass toggle.

Outputs: images and masks (the detected text regions, cropped and masked), json_str (structured detection results - bounding boxes, confidence, per-region text), and text (everything read, as one string).

How to install it

Search comfyui_LLM_party in ComfyUI Manager and install, then restart - no API key setup needed for this specific node. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/heshengtao/comfyui_LLM_party.git

Run pip install -r requirements.txt from inside the pack's own folder using ComfyUI's Python, then restart ComfyUI. EasyOCR itself is a normal Python dependency pulled in by that install; what happens after install - downloading the actual detection/recognition model weights per language - happens automatically the first time you select a given language on the node, not during the pip step.

Common issues & troubleshooting

First run with a new language is slow, or seems to hang. That's EasyOCR downloading the model weights for that language pair in the background - it only happens once per language, and every run after is fast. Give it time on the first pass rather than assuming it's stuck.

Crashes when gpu is true. This means EasyOCR can't find a working CUDA-enabled PyTorch - either you're on a machine without an NVIDIA GPU, or your driver/CUDA/PyTorch versions don't line up. Set gpu to false to force CPU mode; it's slower but has no driver dependency.

Recognition quality is poor on your specific image. EasyOCR does well on printed text in reasonably clean layouts and struggles more with stylized fonts, dense small text, or heavy skew/rotation. If accuracy matters more than speed and you have API budget, the pack's Mini Advanced OCR node (EasyOCR detection plus a vision LLM read-out) or GOT-OCR2 will generally do better on messier inputs - this node trades some of that ceiling for being fully local and free to run.