Author Topic: Ai generated images.  (Read 5746 times)

Offline NancyGold

  • Colonel
  • ****
  • Posts: 186
    • View Profile
Re: Ai generated images.
« Reply #15 on: August 13, 2023, 01:06:02 pm »
Markov models were there since the early 20th century, when Markov himself used 1-gram model to simulate text writing and physical processes. It was already obvious, that by implementing an N-gram model, one will be able to simulate human reasoning or any physical process. So Markov claimed that God doesn't exist, since all nature's processes could ultimately be understood and implemented, which got him excommunicated.

But it took more than a century to properly implement N-gram models, and, believe me, people tried hard ever since computers were invented. There were 2-gram models, 3-gram models, etc... but natural language conversations are many sentences long (i.e. more than 100-gram model is needed), and one needs a lot of training data to for a blind and deaf brain-in-a-vat to understand humans. In fact it required scrapping the entire Internet.

As of now AI is just a tool. You need a lot of expertise to do anything actually useful with it.
1. Understand how foundational models work.
2. Train LoRA.
3. Prompt the model in a consistent way.
4. Detect errors somehow. Model can generate bad art, and unless you're an artist, you wont notice what is precisely wrong.

And it will still look bland and generic, unless you train LoRA on something you personally designed in an unique way.

But in the end it can save you a bit of time or allow to produce photo-realistic content if you really know what you're doing.

Yet in observable future with well trained multimodal modal models it will be a real challenge for humanity to handle, since it could work without anyone guiding it precisely. I.e. a language model will do planning and initial composition, then image model will do a sketch, feeding result back into the language and spatial models, which will fix inconsistencies like 6 fingers.

These image models already have some basic interfacing with language models - they even use the same word embedding scheme. And further multimodal integration is driven by the need in robotics to navigate the complex real world environment.

I'm personally not good at competing even with other people, so guess I have already lost to the AI which will be there in 100 years. So I will just continue working on my projects.
« Last Edit: August 13, 2023, 01:07:36 pm by NancyGold »