How to use pix2struct for pure OCR tasks

[ash2703]

Using the OCR-VQA model does not always give consistent results when the prompt is left unchanged
What is the most consitent way to use the model as an OCR?

[hrithickcodes]

You can use the fine-tuned text-caps model, train it again on ocr tasks which involves learning to generate text from a given image. I hope this will get you there.