I'm still trying to figure that out! AWS Textract is phenomenal at handwriting
@simonw
-
Optimizing Vector Loading: Memory-Efficient Brute Force Filtering
By
–
Right, but you don't have to keep them all in memory for those initial brute force calculations – you can load the 100 floating point vectors from disk once you've filtered down to them
-
GPT-4 Multimodal Training from Original Paper
By
–
It's GPT-4 itself – the GPT-4 model was trained to be multi-modal with images from the start, see the original paper https://
arxiv.org/html/2303.0877
4v6#S4
… -
AI System Struggles with Handwriting Recognition Accuracy
By
–
It's not great at handwriting from my experience
-
Tool Limitation: Processing Long Images Single Frame
By
–
I have a tool I've already built that accepts an image at a time and I was hoping it would work with long images
-
Image Resolution Impact on AI Model Hallucination Risks
By
–
OK, this seems to be key: if the text is illegible it's VERY prone to hallucination. In my case I'd tried turning a PDF into a single long image concatenating pages together, which clearly got resized down to where it could'nt make out individual letters
-
GPT-4 Vision OCR: Prompt Engineering Tips for Better Results
By
–
GPT-4 vision doesn't use tesseract. But it sometimes looks like it does, because if you ask ChatGPT to OCR something it often decides to use PyTesseract instead You need to tell GPT-4 "tell me about the text in this image" without mentioning OCR to avoid using Code Interpreter
-
GPT-4 Vision vs Claude 3 for OCR: Tesseract and Textract comparison
By
–
Just seeing what my options are – I usually use Tesseract locally or AWS Textract for harder stuff, was wondering if GPT-4 Vision or Claude 3 could beat those
-
Gemini Pro 1.5 API Plugin Update for LLM Tool
By
–
For anyone who's on the Gemini Pro 1.5 API preview, I've updated the llm-gemini plugin to support it llm install llm-gemini
llm keys set gemini cat really-long-file.txt | llm -m gemini-1.5-pro-latest "summarize this document" -
GPT-4 Vision Hallucination Issues with Scaled Down Images
By
–
That must be what happened here! I tried turning a PDF into a single long image (concatenating the pages together) and GPT-4 Vision hallucinated most of the document – I bet it got scaled it down to being almost illegible before being fed to the model