deepseek-ai

/

DeepSeek-OCR-2

Like 1.1k

Follow

DeepSeek 148k

Hello friends, a common issue we encounter when running this model is that it can go into an infinite loop of repetition and fail to give a good output.

This happens when the model's vision component and langauge component de-sync. Meaning the language model is no longer following the vision input but going on an adventure of its own.

As a specialist model, DeepSeek-OCR-2 is very senstive to prompting and giving the model a suitable prompt for the document/task is the key way to gain performance.

I've done some analysis on which prompt works best on which type of document.

Some findings:

1. Document <image>\n<|grounding|>Convert the document to markdown. is the best generalist prompt over <image>\nFree OCR.
2. General <image>\nDescribe this image in detail. is the most reliable prompt, though it doesn't usually give you word for word output.
3. Rec <image>\nLocate <|ref|>{OBJECT}<|/ref|> in the image. is effective with photos, but not documents.

You can view the full dataset, including exact model outputs, in an interactive web app here: https://deepseek-ocr-v2-demo.vercel.app

Curious to hear any creative uses of this model + performance tips.