JPEG works in 8x8 pixel blocks, and back in the day, most JPEG images weren’t all that big. Each 8x8 pixel block (64 pixels per block) could easily and quickly be processed as if it were a single pixel.
So if you had a 1024x768 JPEG, then the fast scanning technique would only scan the 128x96 blocks, not necessary to process every single pixel.
Of course the results could never be perfectly accurate, but most images are unique enough that this would be more than sufficient for fast scanning.
Okay, not entirely a layman but also not exactly an expert, if the Photoshop max pixelated entry has the same formula as the detailed comparison it would match? And if that is the case, I imagine all the human input data and behavioral wise would only better the algorithm?
Looking past the days of old, while also dismissing modern artificial intelligence, the same techniques would still work if you just processed the thumbnails of the images, which for simplicity sake, might as well be a 1/8 scale image, if not actually even lower resolution.
That makes sense. Ive seen it do some amazing results but also some painfully hard-to-make mistakes. Minda neat, imagine going by that mindset, making the most with what you have, without a never ending redundant hell of depencies for even the most basic functiin/feature?!
making the most with what you have
That was, indeed, the motto of ML research for a long time. Just hacking out more efficient approaches.
It’s people like Altman that introduced the idea of not innovating and just scaling up what you already have. Hence many in the research community know he’s full of it.
All the other answers are wrong, the real reason is that the browser does black magic by computing the number of pixels divided by zero.
Through reverse osmosis, this can be found to collude to the RGB pixels of the bottom left corner which if the same value can be determined that the color is black.
See the color code for black is #000000 which when turned into RGB is R:00 G:00 B:00. It then reversed the process to count all the white pixels. This can rule out nearly 2/3rds of images and the other ones are outsourced to an AI company.
The company in specific is the company that was just 500 people in India who all shuffle the images on a HUGE table. This is very slow, so the CEO speeds it up by giving them an IV tube filled with 50/50 redbull and radioactive sludge.
Once the AI finds a correct match the images are inserted into a vacuum tube to the basement where buddy the elf is having a dance party with the tough mail people. Due to this, there is only a small chance it will happen to float into the correct output vacuum tube. This is why TinEye is extremely unstable in giving accurate results.
Once the output is given it sends the letter to the local library to photocopy it by the sweet librarian named Edna. Once the image is found it is then e-mailed to HR to be double checked ever since those teenagers scanned their butts that one time.
It is then sent from Edna’s computer to a USB stick when gets stuck in car traffic because of the train festival located in town once every decade.
Once it gets back to TinEye HQ the janitor plugs it into the computer and looks at it even though the IT guy told him not to. The IT guy sees this on the alert system he installed last time and takes it from the janitor to deal with it properly.
The IT guy gives it a ticket number #538,221 on a sticky note then manually uploads it to TinEye.com to get to the customer.
He then closes ticket #538,221.Tap for spoiler
this is a joke answer if you couldn’t tell.
Also I didn’t use AI for this, I’m just bored as hell rn
I did use AI for this though:
.±-------+
.| hello |
.±–v—+
. (_)/
. ( | )
. / \I’m naming him Jeff.
edit: OU NO JEFF FORMATTING KILELD HIM!!!They had the AI models of those days.
That’s cool, didn’t know AI models where a thing in those days. Are they comparable (maybe more crude?) to nowadays tech? Like, did they use machineearning? As far as I remember there were not much dedicated AI accelerating hardware pieces. Maybe a beefy GPU for neural network purposes? Interesting though
Models were a thing even some 30 or 40 years ago. Processing power makes most of the difference today: it allows larger models and quicker results.
I didn’t know. Are you somewhat informed about the history of models? I’d love to hear it from you instead of a random crypto bros’ LLM summary. Thanks!
Machine learning has been a field for years, as others said, yeah, but Wikipedia would be a better expansion of the topic. In a nutshell, it’s largely about predicting outputs based on trained input examples.
It doesn’t have to be text. For example, astronmers use it to find certain kinds of objects in raw data feeds. Object recognition (identifying things in pictures with little bounding boxes) is an old art at this point. Series prediction models are a thing, languagetool uses a tiny model to detect commonly confused words for grammar checking. And yes, image hashing is another, though not entirely machine learning based. IDK what Tineye does in their backend, but there are some more “oldschool” approaches using more traditional programming techniques, generating signatures for images that can be easily compared in a huge database.
You’ve probably run ML models in photo editors, your TV, your phone (voice recognition), desktop video players or something else without even knowing it. They’re tools.
Seperately, image similarity metrics (like lpips or SSIM) that measure the difference between two images as a number (where, say, 1 would be a perfect match and 0 totally unrelated) are common components in machine learning pipelines. These are not usually machine learning based, barring a few execptions like VMAF (which Netflix developed for video).
Text embedding models do the same with text. They are ML models.
LLMs (aka models designed to predict the next ‘word’ in a block of text, one at a time, as we know them) in particular have an interesting history, going back to (If I even remember the name correctly) BERT in Google’s labs. There were also tiny LLMS people did run on personal GPUs before ChatGPT was ever a thing, like the infamous Pygmalion 6B roleplaying bot, a finetune of GPT-J 6B. They were primitive and dumb, but it felt like witchcraft back then (before AI Bros marketers poisoned the well).
As a transmillenial student of AI / ML, great write-up.
I don’t remember too much tbh, just that we heard about the theory at university and tried out some of the mathematical methods. They were tiresome ;)
Today I would recommend to start your studies on the wikipedia pages about Markov models and about machine learning.
Yann Lecun gave us convolutional neural networks (CNNs) in 1998. These are the models that are used for pretty much all specialized computer vision tasks even today. TinyEye came into existence ten years later in 2008. I can’t tell you if they used CNNs, but they were certainly available.
Oh and to answer this, specifically, Nvidia has been used in ML research forever. It goes back to 2008 and stuff like the desktop GTX 280/CUDA 1.0. Maybe earlier.
Most “AI accelerators” are basically the same thing these days: overgrown desktop GPUs. They have pixel shaders, ROPs, video encoders and everything, with the one partial exception being the AMD MI300X and beyond (which are missing ROPs).
CPUs were used, too. In fact, Intel made specific server SKUs for giant AI users like Facebook. See: https://www.servethehome.com/facebook-introduces-next-gen-cooper-lake-intel-xeon-platforms/
We didn’t call them AI because they weren’t (and aren’t) intelligent, but marketing companies eventually realized there were trillions of dollars to be made convincing people they were intelligent and created models explicitly designed to convince people of things like the idea that they are intelligent and can have genuine conversations like a real human and create real art like a real human and totally aren’t just empty-headedly mimicking thousands of years of human conversation and art, and immediately used them to convince people that the models themselves were intelligent (and many other things besides). Given that marketing and advertising literally exist to convince people of various things and have become exceedingly good at it, it’s really a brilliant business move and seems to be working great for them.
We didn’t call them AI because they weren’t (and aren’t) intelligent, but marketing companies eventually realized there were trillions of dollars to be made convincing people they were intelligent and created models explicitly designed to convince people of things like the idea that they are intelligent and can have genuine conversations like a real human and create real art like a real human and totally aren’t just empty-headedly mimicking thousands of years of human conversation and art, and immediately used them to convince people that the models themselves were intelligent (and many other things besides). Given that marketing and advertising literally exist to convince people of various things and have become exceedingly good at it, it’s really a brilliant business move and seems to be working great for them.
https://vincmazet.github.io/bip/filtering/fourier.html
There are ways to encode images that make it easier to isolate differences in cropping and resolution and rotation. Like how if you wanted to search for color filtered images you could just throw out the color and compare them in black and white.
What you’re looking for is the history of “computer vision”
Looking up similar images and searching for crops are computer vision topics, not large language model (basically text predictor) or image generation ai topics.
Image hashing has been around for quite a while now and there is crop resistant image hashing libraries readily available like this one: https://pypi.org/project/ImageHash/
It’s basically looking for defining features in images and storing those in an efficient searchable way probably in a traditional database. As long as they are close enough or in the case of a crop, a partial match, it’s a similar image.
