Easy Guide to How AI Generates Images

In this video I am offerring up a non-technical explaination of how AI takes your text input and generates images from your specification and description to provide the highest quality imagery that is possible with current Generatice AI models. These models are created ahead of time, but are continously refined overtime by user interaction, and choices users make when picking just the right image for them. I use the pronounciation JAN for GANs because GAN is an overloaded IT term, even though the previous use is obsolete there are some of us old guys still around who might confuse GANs. In case you are wondering...a GAN used to mean Global Area Network, GaN used to mean Gallium Nitride and was a type of Semi-conductor.

AI Generated Thumbnail: based on my real image

00:00 - Intro
00:14 - Step 1: The Front End
01:04 - How AI is Trained to Generate Images
02:40 - Refining the AI Model
03:53 - Ai generated Images
05:49 - AI Model Generation for Images
06:08 - How AI Evaluates Image Quality
08:47 - Neural Networks
09:23 - Human Evaluations
10:00 - Consistency Checks
12:07 - AI Image Creation
15:18 - Advanced Techniques
16:22 - AI Replacing Creatives?
19:07 - Conclusions

Follow me:
Twitter @djware55

[Scranny]

DJ I much prefer watching you talk to the camera than all these short stock clips interleaved in your video. :-)

[carpetbomberz]

This is just the right level of detail, 💯, thanks for weighing in, definitely worthwhile.

[paulhiggins]

Human Artists learn anatomy, colour theory, composition, lighting, perspective, shape language and many other subtle and complex things- and it is from their understanding of these principles that artists develop their ability to render three dimensional reality in the form of two dimensional shapes and colours.

AI Art generators know absolutely nothing about any of these things and have no capability to apply them. All AI Art generators do is apply statistical math to vast data sets of linked images and word pairs.

However, because the data sets they are trained on contain the accumulated knowlege and skill of millions of human artists it's easy to become confused and assume that the AI knows what it is doing.

The best analogy here is to imagine a parrot that has been trained to respond to the name 'Einstein' by reciting the phrase 'E= M, C, squared' The uninformed observer might conclude that here we have a parrot that understands special relativity- but on closer inspection discovers that the parrot is just a parrot and has no real idea what it is talking about- it has simply learned to create a certain pattern of sound in response to another pattern of sound.

There really is no such thing as 'Artificial Intelligence' because to build such a thing we would first have to understand 'non artificial intelligence' and the truth is we barely understand our own intelligence- so the idea that we have now- or might in the future- create artificial versions of our own intelligence is pure science fiction.

AI is mostly hype with a dash of genuine innovation thrown in- but anyone expecting truly intelligent machines to arrive any time soon is going to be really disappointed when they find out that the parrot has no real understanding of what the noises it makes actually mean.

Also it's worth observing that Artists don't talk to their tools, they use them directly in a non verbal way to translate the images in their minds to images on the canvas or screen. Using words to define images is a really stupid interface design for the obvious reason that if words really could convey images in a precise way you would not need the images- the words alone would be enough.

If I paid the entry cost to see the Mona Lisa in Paris and was instead presented with a sheet of paper upon which the words 'The Mona Lisa' was printed- followed by a detailed description of the painting I would not be happy because reading about the painting is not the same as seeing it.

The truth is that even the most complete text description of a face or a scene barely scratches the surface of the true visual complexity that is there- this is why no one will ever succeed in creating a self portrait using a text to image interface- how would you even begin to describe to the AI in words the subtle complexity of your own face in enough detail to allow that AI to produce an accurate image? It's simply not possible.

So the idea of 'text to image' as a serious way to create visual art is logically nonsense- if words and pictures were interchangeable then all the blind would need to see is a device that described in words the world around them- but we all know that 'seeing' is far more than that. The AI artist is not 'creating' art- they are simply activating a slot machine that responds to word combinations by producing image combinations- with a random seed value thrown in to ensure that a degree of randomness in the outputs creates the illusion that something 'creative' is going on.

As a Kid i was facinated by Kaleidoscopes- the way they created a seemingly endless series of patterns and designs- but no one would try to claim that these devices create art- would they?

[eugenesmirnov]

Seems a long way to R'Daniel O'livo
So while we're waiting Ollama podcast will be nice.