Invoice Reader with Small VLM

Basic Web application setup with Flask and Small Vision Language Model It is a very straight-forward technique of a web app with a file upload, after submitting the VLM inference is called to parse the image file and describe it.

Tools & Libraries

• Python 3.10
• Flask==3.0.3
• torch==2.6.0
• transformers
• pillow==11.1.0
• python-docx==1.1.2

Model

SmolVLM-500M-Instruct model by HuggingfaceTB is used for inferring, as it is an Image+Text to Text small language model. The model size is approximtely around 1.1GB and it is downloaded only once.

NOTE : This is a basic setup for the image parsing using VLM. It can be improved further more.