comfyui_glm_ocr

ComfyUI custom node to run GLM-OCR locally. Loads the model straight from your ComfyUI models folder — no Ollama, no API calls.

---

What it does

GLM-OCR is a 0.9B vision-language model built for document understanding. You give it an image and a task prompt, it gives you text back. Works well on:

• Scanned documents and photos of text
• Tables (outputs markdown)
• Math formulas (outputs LaTeX)
• General image description

---

Nodes

GLM-OCR Model Loader

Scans every checkpoint directory ComfyUI knows about (including paths from extra_model_paths.yaml) and lists any subfolder that contains a config.json. Select your model from the dropdown — no hardcoded paths.

GLM-OCR Run

Takes the loaded model + an image and runs inference.

| Input | Description | |-------|-------------| | glm_ocr_model | from the loader node | | image | any ComfyUI IMAGE | | task_prompt | preset task — see table below | | max_new_tokens | max output length (default 2048) | | custom_prompt | leave empty to use task_prompt, or type anything |

| Task Prompt | Use | |-------------|-----| | Text Recognition: | extract all text from the image | | Formula Recognition: | extract math as LaTeX | | Table Recognition: | extract tables as markdown | | Describe the image. | general description |

Output is a STRING — connect it to ShowText or any other text node.

---

Installation

1. Install the node

Clone into your ComfyUI/custom_nodes/ folder:

cd ComfyUI/custom_nodes
git clone https://github.com/Dangocan/comfyui_glm_ocr

Or use ComfyUI Manager → Install via Git URL → paste the repo URL.

2. Download GLM-OCR

Put the model in any checkpoint directory ComfyUI knows about. Example:

ComfyUI/models/checkpoints/GLM-OCR/

The folder needs to contain config.json, model.safetensors, and the tokenizer files. Download with:

huggingface-cli download zai-org/GLM-OCR --local-dir ComfyUI/models/checkpoints/GLM-OCR

Or download manually from https://huggingface.co/zai-org/GLM-OCR/tree/main

3. Check your transformers version

GLM-OCR requires transformers >= 5.3.0. Run this to check:

python -c "import transformers; print(transformers.__version__)"

If it's below 5.3.0 or you get an error about glm_ocr architecture not being recognized, install from source:

pip install git+https://github.com/huggingface/transformers.git

4. Restart ComfyUI

The nodes show up under the GLM-OCR category.

---

Extra model paths

If your models are on a separate drive, add it to extra_model_paths.yaml in your ComfyUI root:

comfyui:
    base_path: D:/MyModels/
    checkpoints: checkpoints/

The loader node picks up any registered checkpoint path automatically.

---

Requirements

• ComfyUI
• Python 3.10+
• PyTorch (CUDA recommended)
• transformers >= 5.3.0 (install from git — see above)
• Pillow

---

Model

zai-org/GLM-OCR — https://huggingface.co/zai-org/GLM-OCR

MIT License