The crucial information is missing - accuracy comparison with other OCR providers. From my experience LLM based OCR might misread the layout and hallucinate values, it is very subtle but sometimes critically wrong. Classical OCR has more precision but doesn't get the layout at all. Combining both has other issues, no approach is 100% reliable.
Have you evaluated this lately? Last year or even just earlier this year I would have mostly agreed with you. At this point, however, with at least the documents I have been working on, OCR reliability with GPT5 or Mistral OCR [1] has been much better than even domain-trained classical OCR. If the documents have even slightly complex layout (to say nothing of page numbers or page headings or an uncommon font), the accuracy of state of the art LLMs has been in my work an order of magnitude greater. The ability to have the LLM tentatively combine trailing sentences across pages, which is especially useful if you have to work with documents in German say, is invaluable.
I asked GPT-5 to OCR a table for me the other day, it hallucinated perhaps 10% of the values. This was a screenshot of a spreadsheet, with large font, not challenging except for the layout.
What's interesting is that I asked it to also read the background colors of the cells and it did much worse on that task.
I believe these models could be useful for a first pass if you are willing to manually review everything they output, but the failure mode is unsettling.
Breaking up the page, feeding the pieces one-by-one and reassembling the output helps with that. I was expecting this project to do that but it can only feed a whole page.
By 1990 Omnipage 3 and its successors were 'good enough' and with their compact dictionaries and letter form recognition were miracles of their time at ~300MB installed.
In 2025 LLMs can 'fake it' using Trilobites of memory and Petaflops. It's funny actually, like a supercomputer being emulated in real time on a really fast Jacquard loom. By 2027 even simple hand held calculator addition will be billed in kilowatt-hours.
I had an on-screen OCR app on my Amiga in the early 90s that was amazing, so long as the captured text image used a system font. Avoiding all the mess of reality like optics, perspective, sensors and physics and it could be basically perfect.
Tesseract can do wonders for scanned paper (and web generated PDFs) both in its old and new version. If you want to pay for something closed, Prizmo on macOS is extremely good as well.
On the other hând, LLm5 are sl0wwer, moré resource hangry and l3ss accurale fr their outpu1z.
A bit ago I tried throwing a couple of random simple Japanese comics (think 4koma but I don't think either of the ones I threw in were actually 4 panels) from Pixiv into Gemma 3b on AI studio.
- It transcribed all of the text, including speech, labels on objects, onomatopoeias in actions, etc. I did notice a kana was missing a diacritic in a transcription, so the transcriptions were not perfect, but pretty close actually. To my eye all of the kanji looked right. Latin characters already OCR pretty well, but at least in my experience other languages can be a struggle.
- It also, unprompted, correctly translated the fairly simple Japanese to English. I'm not an expert, but the translations looked good to me. Gemini 2.5 did the same, and while it had a slightly different translation, both of them were functionally identical, and similar to Google Translate.
- It also explained the jokes, the onomatopoeias, etc. To my ability to verify these things they seemed to be correct, though notably Japanese onomatopoeias used for actions in comics is pretty diverse and not necessarily super well-documented. But contextually it seemed right.
To me this is interesting. I don't want to anthropomorphize the models (at least unduly, though I am describing the models as if they chose to do these things, since it's natural to do so) but the fact that even relatively small local models such as Gemma can perform tasks like this on arbitrary images with handwritten Japanese text bodes well. Traditional OCR struggles to find and recognize text that isn't English or is stylized/hand-written, and can't use context clues or its own "understanding" to fill in blanks where things are otherwise unreadable; at best they can take advantage of more basic statistics, which can take you quite far but won't get you to the same level of proficiency at the job as a human. vLLMs however definitely have an advantage in the amount of knowledge embedded within them, and can use that knowledge to cut through ambiguity. I believe this gets them closer.
I've messed around with using vLLMs for OCR tasks a few times primarily because I'm honestly just not very impressed with more traditional options like Tesseract, which sometimes need a lot of help even just to find the text you want to transcribe, depending on how ideal the case is.
On the scale of AI hype bullshit, the use case of image recognition and transcription is damn near zero. It really is actually useful here. Some studies have shown that vLLMs are "blind" in some ways (in that they can be made to fail by tricking them, like Photoshopping a cat to have an extra leg and asking how many legs the animal in the photo has; in this case the priors of the model from its training data work against it) and there are some other limitations (I think generally when you use AI for transcription it's hard to get spatial information about what is being recognized, though I think some techniques have been applied, like recursively cutting an image up and feeding it to try to refine bounding boxes) but the degree to which it works is, in my honest opinion, very impressive and very useful already.
I don't think that this demonstrates that basic PDF transcription, especially of cleanly-scanned documents, really needs large ML models... But on the other hand, large ML models can handle both easy and hard tasks here pretty well if you are working within their limitations.
Personally, I look forward to seeing more work done on this sort of thing. If it becomes reliable enough, it will be absurdly useful for both accessibility and breaking down language barriers; machine translation has traditionally been a bit limited in how well it can work on images, but I've found Gemini, and surprisingly often even Gemma, can make easy work of these tasks.
I agree these models are inefficient, I mean traditional OCR aside, our brains do similar tasks but burn less electricity and ostensibly need less training data (at least certainly less text) to do it. It certainly must be physically possible to make more efficient machines that can do these tasks with similar fidelity to what we have now.
100%. My sense is that many in this thread have never gone through the misery of trying to use classical OCR for non-English documents or where you can't control scan quality. I did a test recently with 18th-century German documents, written in a well-known and standardized but archaic script. The accuracy of classical models specifically trained on this corpus was an order of magnitude lower than GPT5. I haven't experimented personally or professionally with smaller models, but your experience makes me hopeful that we might even get this accurate OCR on phones sooner rather than later...
- Optionally extract images/graphs within the page and include them in the same LLM call.
- Optionally add a bit of context from neighboring pages.
Then wrap everything with a clear prompt (structured output + how you want graphs handled), and you’re set.
At this point, models like GPT-5-nano/mini or Gemini 2.5 Flash are cheap and strong enough to make this practical.
Yeah, it’s a bit like using a rocket launcher on a mosquito, but this is actually very easy to implement and quite flexible and powerfuL. works across almost any format, Markdown is both AI and human friendly, and surprisingly maintainable.
You don't need full reasoning to get accurate results, so even with GPT5 it's still pretty cheap for a one-time job and easy to reason about costs. It's certainly cheaper if you have data where reliability is key and classical OCR will undoubtedly require some manual data cleaning...
I can recommend the Mistral OCR API [1] if you have large jobs and don't want to think about it too much.
I really wanted this to be good. Unfortunately it converted a page that contained a table that is usually very hard for converters to properly convert and I got a full page with "! Picture 1:" and nothing else. On top of that, it hung at page 17 of a 25 page document and never resumed.
As far as I am aware, the "hanging" issue remains unsolved to this day. The underlying problem is that LLMs sometimes get stuck in a loop where they repeat the same text again and again until they reach the token limit. You can break the loop by setting a repeat penalty, but when your image contains repeated text, such as in tables, the LLM will output incorrect results to prevent repetition.
Here is the corresponding GitHub issue for your default model (Qwen2.5-VL):
You can mitigate the fallout of this repetition issue to some degree by chopping up each page into smaller pieces (paragraphs, tables, images, etc.) with a page layout model. Then at least only part of the text is broken instead of the entire page.
A better solution might be to train a model to estimate a heat map of character density for a page of text. Then, condition the vision-language model on character density by feeding the density to the vision encoder. Also output character coordinates, which can be used with the heat map to adjust token probabilities.
I’ve been trying to convert a dense 60 page paper document to Markdown today from photos taken on my iPhone. I know this is probably not the best way to do it but it’s still been surprising to find that even the latest cloud models are struggling to process many of the pages. Lots of hallucination and “I can’t see the text” (when the photo is perfectly clear). Lots of retrying different models, switching between LLMs and old fashioned OCR, reading and correcting mistakes myself. It’s still faster than doing the whole transcription manually but I thought the tech was further along.
Looking at the code, this converts PDF pages to images, then transcribes each image. I might have expected a pdftotext post-processor. The complexity of PDF I guess ...
20 years ago I tried in vain to get my HOA to use the virtual printer for PDF documents so they'd be searchable. The capability was built in to both Mac and Windows even way back then.
No luck. They just could not grasp it. So they kept using their process of printing out the file on paper and then scanning it back in as a PDF image file.
I finally quit trying. Now of course they've seen the light and are painstakingly OCRing all that old stuff.
> annotateai automatically annotates papers using Large Language Models (LLMs). While LLMs can summarize papers, search papers and build generative text about papers, this project focuses on providing human readers with context as they read.
Image-based extraction often preserves layout and handles PDFs with embedded fonts, scanned content, or security restrictions better than direct text extraction methods.
Readit News
[1] https://mistral.ai/news/mistral-ocr
What's interesting is that I asked it to also read the background colors of the cells and it did much worse on that task.
I believe these models could be useful for a first pass if you are willing to manually review everything they output, but the failure mode is unsettling.
>OCRmyPDF adds an OCR text layer to scanned PDF files, allowing them to be searched
I ... I nailed it.
In 2025 LLMs can 'fake it' using Trilobites of memory and Petaflops. It's funny actually, like a supercomputer being emulated in real time on a really fast Jacquard loom. By 2027 even simple hand held calculator addition will be billed in kilowatt-hours.
On the other hând, LLm5 are sl0wwer, moré resource hangry and l3ss accurale fr their outpu1z.
We shoulD stop gl0rıfying LLMs for 3verylhin9.
Dead Comment
Trilobites? Those were truly primitve computers.
- It transcribed all of the text, including speech, labels on objects, onomatopoeias in actions, etc. I did notice a kana was missing a diacritic in a transcription, so the transcriptions were not perfect, but pretty close actually. To my eye all of the kanji looked right. Latin characters already OCR pretty well, but at least in my experience other languages can be a struggle.
- It also, unprompted, correctly translated the fairly simple Japanese to English. I'm not an expert, but the translations looked good to me. Gemini 2.5 did the same, and while it had a slightly different translation, both of them were functionally identical, and similar to Google Translate.
- It also explained the jokes, the onomatopoeias, etc. To my ability to verify these things they seemed to be correct, though notably Japanese onomatopoeias used for actions in comics is pretty diverse and not necessarily super well-documented. But contextually it seemed right.
To me this is interesting. I don't want to anthropomorphize the models (at least unduly, though I am describing the models as if they chose to do these things, since it's natural to do so) but the fact that even relatively small local models such as Gemma can perform tasks like this on arbitrary images with handwritten Japanese text bodes well. Traditional OCR struggles to find and recognize text that isn't English or is stylized/hand-written, and can't use context clues or its own "understanding" to fill in blanks where things are otherwise unreadable; at best they can take advantage of more basic statistics, which can take you quite far but won't get you to the same level of proficiency at the job as a human. vLLMs however definitely have an advantage in the amount of knowledge embedded within them, and can use that knowledge to cut through ambiguity. I believe this gets them closer.
I've messed around with using vLLMs for OCR tasks a few times primarily because I'm honestly just not very impressed with more traditional options like Tesseract, which sometimes need a lot of help even just to find the text you want to transcribe, depending on how ideal the case is.
On the scale of AI hype bullshit, the use case of image recognition and transcription is damn near zero. It really is actually useful here. Some studies have shown that vLLMs are "blind" in some ways (in that they can be made to fail by tricking them, like Photoshopping a cat to have an extra leg and asking how many legs the animal in the photo has; in this case the priors of the model from its training data work against it) and there are some other limitations (I think generally when you use AI for transcription it's hard to get spatial information about what is being recognized, though I think some techniques have been applied, like recursively cutting an image up and feeding it to try to refine bounding boxes) but the degree to which it works is, in my honest opinion, very impressive and very useful already.
I don't think that this demonstrates that basic PDF transcription, especially of cleanly-scanned documents, really needs large ML models... But on the other hand, large ML models can handle both easy and hard tasks here pretty well if you are working within their limitations.
Personally, I look forward to seeing more work done on this sort of thing. If it becomes reliable enough, it will be absurdly useful for both accessibility and breaking down language barriers; machine translation has traditionally been a bit limited in how well it can work on images, but I've found Gemini, and surprisingly often even Gemma, can make easy work of these tasks.
I agree these models are inefficient, I mean traditional OCR aside, our brains do similar tasks but burn less electricity and ostensibly need less training data (at least certainly less text) to do it. It certainly must be physically possible to make more efficient machines that can do these tasks with similar fidelity to what we have now.
For each page:
- Extract text as usual.
- Capture the whole page as an image (~200 DPI).
- Optionally extract images/graphs within the page and include them in the same LLM call.
- Optionally add a bit of context from neighboring pages.
Then wrap everything with a clear prompt (structured output + how you want graphs handled), and you’re set.
At this point, models like GPT-5-nano/mini or Gemini 2.5 Flash are cheap and strong enough to make this practical.
Yeah, it’s a bit like using a rocket launcher on a mosquito, but this is actually very easy to implement and quite flexible and powerfuL. works across almost any format, Markdown is both AI and human friendly, and surprisingly maintainable.
It all depends on the scale you need them, with the API it's easy to generate millions of tokens without thinking.
I can recommend the Mistral OCR API [1] if you have large jobs and don't want to think about it too much.
[1] https://mistral.ai/solutions/document-ai
Here is the corresponding GitHub issue for your default model (Qwen2.5-VL):
https://github.com/QwenLM/Qwen2.5-VL/issues/241
You can mitigate the fallout of this repetition issue to some degree by chopping up each page into smaller pieces (paragraphs, tables, images, etc.) with a page layout model. Then at least only part of the text is broken instead of the entire page.
A better solution might be to train a model to estimate a heat map of character density for a page of text. Then, condition the vision-language model on character density by feeding the density to the vision encoder. Also output character coordinates, which can be used with the heat map to adjust token probabilities.
I’d just prefer that any images and diagrams are copied over, and rendered into a popular format like markdown.
https://github.com/rednote-hilab/dots.ocr
Seems to weigh about 6GB which feels reasonable to manage locally
https://github.com/ocrmypdf/OCRmyPDF
No LLMs required.
No luck. They just could not grasp it. So they kept using their process of printing out the file on paper and then scanning it back in as a PDF image file.
I finally quit trying. Now of course they've seen the light and are painstakingly OCRing all that old stuff.
Python: PyPdf2, PdfMiner.six, Grobid, PyMuPdf; pytesseract (C++)
paperetl is built on grobid: https://github.com/neuml/paperetl
annotateai: https://github.com/neuml/annotateai :
> annotateai automatically annotates papers using Large Language Models (LLMs). While LLMs can summarize papers, search papers and build generative text about papers, this project focuses on providing human readers with context as they read.
pdf.js-hypothes.is: https://github.com/hypothesis/pdf.js-hypothes.is:
> This is a copy of Mozilla's PDF.js viewer with Hypothesis annotation tools added
Hypothesis is built on the W3C Web Annotations spec.
dokieli implements W3C Web Annotations and many other Linked Data Specs: https://github.com/dokieli/dokieli :
> Implements versioning and has the notion of immutable resources.
> Embedding data blocks, e.g., Turtle, N-Triples, JSON-LD, TriG (Nanopublications).
A dokieli document interface to LLMs would be basically the anti-PDF.
Rust crates: rayon handles parallel processing, pdf-rs, tesseract (C++)
pdf-rs examples/src/bin/extract_page.rs: https://github.com/pdf-rs/pdf/blob/master/examples/src/bin/e...
https://threadreaderapp.com/thread/1955355127818358929.html