Post B4W7DiruMr93GPLSYi by chjara@akko.tuxcrafting.xyz
(DIR) More posts by chjara@akko.tuxcrafting.xyz
(DIR) Post #B4W6Dol8MlFnZyer9U by jessebot@social.smallhack.org
1 likes, 2 repeats
Hey Linux friends, what's your favorite OCR tool that doesn't use LLMs? Bonus points if it works from the command line :)#linux #ocr #noai #debian
(DIR) Post #B4W6KAXyGypgtFWuKu by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@jessebot@social.smallhack.org Wait, which OCR tools use LLMs?
(DIR) Post #B4W6LhXD6vfhUOu8q8 by fiore@brain.worm.pink
0 likes, 0 repeats
@jessebot whats ocr ?
(DIR) Post #B4W75ogypbFp0cSoUq by fiore@brain.worm.pink
0 likes, 0 repeats
@jessebot why are u boosting me im asking a genuine question 😠ive jeard the term before but its slipping my mind rn im curious
(DIR) Post #B4W7DiruMr93GPLSYi by chjara@akko.tuxcrafting.xyz
0 likes, 0 repeats
@fiore @jessebot Optical Character Recognition.
(DIR) Post #B4W7Dj65W8UHyNymf2 by fiore@brain.worm.pink
0 likes, 0 repeats
@chjara @jessebot thank u ..
(DIR) Post #B4W7Gg79434JaesiDQ by jessebot@social.smallhack.org
0 likes, 0 repeats
@fiore sorry, didn't meant to boost. I meant to say it stands for "optical character recognition" which basically means getting text from other file formats. In my case, I was looking for it from images.This is the project I've used to do it:https://github.com/tesseract-ocr/tesseract?tab=readme-ov-file#installing-tesseractSo you can do:# this can be whatever your distro's package manager issudo apt install tesseract# this will transcribe the imagetesseract Screenshot_20260322_141003.png thing# then you can just cat out the file it createscat thing.txtHave fun!
(DIR) Post #B4W7GgPDypWwUjL9Oa by fiore@brain.worm.pink
0 likes, 0 repeats
@jessebot yay thx
(DIR) Post #B4W7HAJamKqnFww4mW by meeper@udongein.xyz
0 likes, 0 repeats
@jessebot tesseract?
(DIR) Post #B4W7MwMpY69rjfCiX2 by meeper@udongein.xyz
0 likes, 0 repeats
@SuperDicq @jessebot I mean you can spin a text-visual model using ollamaLLMs have some advantage as they *can* make sense of some semantic detail
(DIR) Post #B4W7P3DT3VVTYYpacq by jessebot@social.smallhack.org
0 likes, 0 repeats
@fiore I accidentally clicked the button. Sorry! :[ I unboosted!
(DIR) Post #B4W7P3Si8phSJpxlNw by fiore@brain.worm.pink
0 likes, 0 repeats
@jessebot nono dw haha thanks
(DIR) Post #B4W7VeWlFKxb0q0Gyu by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@meeper@udongein.xyz @jessebot@social.smallhack.org But that's not an LLM.
(DIR) Post #B4W7dZzS02eXG0CcMq by meeper@udongein.xyz
0 likes, 0 repeats
@SuperDicq @jessebot I mean recent llms that supports image input not simple deep learningmultimodal or whatever word they use
(DIR) Post #B4W8H2FerGzICab6Dg by jessebot@social.smallhack.org
0 likes, 0 repeats
@SuperDicq So according to this, at least ONLYOFFICE uses it: https://www.tecmint.com/best-linux-ocr-tools/But also, I came from macOS and recently installed Debian Bananas. MacOS has a bunch of AI baked into all sorts of stuff, the most common way to do OCR there is open an image in the Preview app, which will then use an LLM to get your text. However, that doesn't matter anymore, because I'm not using macOS full time anymore (I still have to at work sometimes, but I'm no longer running it at home).I just wanted to make sure I didn't get a bunch of AI bros in my mentions, so I always mention not to recommend slop when asking for suggestions :)
(DIR) Post #B4W8H2UtwbBGxrjGym by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@jessebot@social.smallhack.org You're reading slop articles and marketing material.LLMs are "Large Language Models". They can not interpret images, they can only work with text. All OCRs (Optical Character Recognition) is also often done using AI models (usually very small ones), but it is never using LLMs, because LLMs do not take images as input.
(DIR) Post #B4W8dYybrsQK4GMi6C by jessebot@social.smallhack.org
0 likes, 0 repeats
@SuperDicq Sure, I went ahead and updated my post to reflect that I also hate AI, so it's clear. Thanks!
(DIR) Post #B4W8dZECvstsqdfAPY by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@jessebot@social.smallhack.org I saw you went with Tesseract. This software also uses AI models.There's nothing wrong or unethical about their AI model, but it is an AI model nonetheless. So if you hate AI you probably don't want to use it (or any OCR software at all).
(DIR) Post #B4W9nDuf8YDdadgnI0 by meeper@udongein.xyz
0 likes, 0 repeats
@SuperDicqthere is nothing stopping you from training a transformer to take text and image input or take image input and output text and use similar methods of training, its like your image editing diffusion models but they output text.and if you have a text input you can simply work them like a usual lbullshit generators.llm doesn't mean a bullshit generating chatbot however and a pure simple ocr but someowh trained on text samples and capable of capturing the same semantic relationships (ie *modelling* language) is a language model.Again potentially being able to use semantic information during ocr instead of brute-forcing those rleationships by solely usinf a cnn is probably capable of good results
(DIR) Post #B4W9x30OXtf4em3TH6 by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@meeper@udongein.xyz If you take a LLM and transform it to no longer work with language doesn't that no longer make it an LLM?
(DIR) Post #B4WAWYRRes5wI2deVs by meeper@udongein.xyz
0 likes, 0 repeats
@SuperDicq ocr is working with language just in image form, it doesnr mean textlanguage modelling simply means that in some sense of the word it has captured the syntax of a language, which can be useful in ocr if such details are relevant befoee coversion to pure text.Again nothing stopping simply creating a oure bullshit genrator llm with extra ports for image information (probably behind some laryers of a cnn like makeup) and you get both a bullshit generator but one that can *see* an image though ofc that doesn't mean you need to architecture iand train it like a bs generator chatbot simoly for ocr
(DIR) Post #B4WAdDUZwizzLMHUEi by meeper@udongein.xyz
0 likes, 0 repeats
@SuperDicq even models for purely translation are language models as they capture language details, don't be mistaken that language modellingmeans bullshit genration
(DIR) Post #B4WBBUGY5XPjDD3mtM by karebu@social.karebu.gay
1 likes, 0 repeats
@SuperDicq @jessebot LLMs have been able to interpret images and even audio for a while nowthere are already many OCR projects that make use of small models specifically trained for the purpose of text and table recognition github.com/PADDLEPADDLE/PADDLEOCR github.com/Tencent-Hunyuan/HunyuanOCR github.com/zai-org/GLM-OCR
(DIR) Post #B4WBBUYH1dam6BLwWG by SuperDicq@minidisc.tokyo
0 likes, 0 repeats
@karebu@social.karebu.gay @jessebot@social.smallhack.org I guess the definition of LLM is not that clear. Because personally I'd say that if you take a LLM and transform it in such a way that it takes images as input instead it is not really an LLM anymore, even though the training method is similar.