ComfyUI-LightOnOCR-2

ComfyUI nodes for LightOnOCR-2-1B, LightOn's 1B-parameter OCR vision-language model. Extracts clean text from documents, receipts, tables, and more.

Installation

1. Clone or copy this repo into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/MennilTossFlykune/ComfyUI-LightOnOCR-2.git

2. Install the required dependencies (from your ComfyUI Python environment):

pip install "transformers>=5.0.0" pillow

3. Restart ComfyUI.

Nodes

Both nodes appear under the LightOnOCR-2 category.

LightOnOCR-2 Model Loader

Loads the model and processor. Connect its output to the Run node.

Input	Type	Description
model_name	combo	Model variant to load (default: lightonai/LightOnOCR-2-1B)
device	combo	auto, cuda, or cpu
dtype	combo	auto, bfloat16, float16, or float32

LightOnOCR-2 Run

Runs OCR on an image and outputs the extracted text.

Input	Type	Description
model	LIGHTONOCR2_MODEL	Output from the Model Loader node
image	IMAGE	Any ComfyUI image
max_tokens	INT	Maximum output tokens (default: 1024, range: 64-8192)
seed	INT	Random seed for reproducibility (default: 42)
do_sample	BOOLEAN	Enable sampling; when off, uses greedy decoding (default: True)
temperature	FLOAT	Sampling temperature (default: 0.2, range: 0.01-2.0)
top_p	FLOAT	Top-p / nucleus sampling threshold (default: 0.9, range: 0.0-1.0)

Output	Type
text	STRING

Supported Model Variants

• lightonai/LightOnOCR-2-1B — Best OCR model
• lightonai/LightOnOCR-2-1B-base — Base model for fine-tuning
• lightonai/LightOnOCR-2-1B-bbox — OCR with image bounding boxes
• lightonai/LightOnOCR-2-1B-bbox-base — Base bbox model for fine-tuning
• lightonai/LightOnOCR-2-1B-ocr-soup — Merged variant for extra robustness
• lightonai/LightOnOCR-2-1B-bbox-soup — Merged OCR + bbox variant