• Quick note - the problem with Youtube videos not embedding on the forum appears to have been fixed, thanks to ZiprHead. If you do still see problems let me know.

ChatGPT

Checkmite said:
The companies developing these AI image generators are doing it for profit.

Commercial use is commercial use; they should have gotten permission, and paid as appropriate, for any material they actively used to develop their product.

Yes, a person can generally download any given image from the internet for personal noncommercial purposes. That is widely recognized to be a different circumstance and isn't relevant.
Click to expand...

Copyright law is different throughout the world so there is no single answer that will cover all the laws but overall the highlighted above is a common misconception. If you are using someone's copyrighted work it makes not an iota of a difference in regard to the legality of using that image whether you are using it for commercial or personal use.

In the USA the only differnce would be in someone trying to gain restitution for your use of their copyrighted work, they ain't going to get much from Lilly who used the copyrighted front cover image of a book in her book report, but McDonalds printing your copyrighted image on every big Mac wrapper.....

There are a few exceptions to being able to use a copyrighted image or photograph without permission, for example fair use (which by the way is not the same in all countries) or for research.

And that research exception is what the creators of many of the datasets collated from publicly accessible websites have relied on for many years and it went unchallenged until very recently. The claim of those using generative AI in commercial projects is that since the generated AI does not store the copyrighted images and does not use them to generate new images then there is no breach of copyright even though they used the copyrighted image to train the AI i.e. they claim that is covered by the research exception.

It is going to take either new legislation or case law to settle this issue.
 
Last edited:
Olmstead said:
Some sort of licensing law?

If the material were software, I couldn't get away with making it part of my code, no matter how different the final application is. I don't see why non-software stuff should be treated any differently.

Is there a lawyer on the plane?





Speaking of false analogies, show me exactly how what a human brain is doing is analogous to what the current crop of "AIs" are doing. Show me the code of the brain that explicitly contains all those copyrighted datasets. AI developers are full of ****. The copyrighted material is in there, in the software, no matter how much it is buried. And the AI keeps referencing those datasets. If there isn't a law against this, there should be.
Click to expand...


The AIs also don't contain the copyrighted datasets.
 
Darat said:
The AIs also don't contain the copyrighted datasets.
Click to expand...

Really? Gee, I wonder why they need all that stuff then. I guess it's turned into data magic, and the AI just magically knows how to ape someone's style. There's nothing in the code at all.
 
Olmstead said:
Really? Gee, I wonder why they need all that stuff then. I guess it's turned into data magic, and the AI just magically knows how to ape someone's style. There's nothing in the code at all.
Click to expand...

Indeed there isn't. Especially not in the code. The model (not the code) contains abstraction of training dataset.

For example Stable Diffussion model (version 1.5) has 2 gigabytes. The training set was 160 millions of 512x512 images. That's compression of about 65000:1.

The abstraction is good though. In many aspects better than what people are capable. It can create photorealistic images at ease. It can recall style of hundreds of artists. If you can identify it exactly enough, it can recall specific training image quite well. Like my example with Mona Lisa.

So in general, you are right, but please allow us to be nitpicky. The pictures are not stored in the code. They are in the model, and they are stored very loosely.
 
Last edited:
Olmstead said:
Really? Gee, I wonder why they need all that stuff then. I guess it's turned into data magic, and the AI just magically knows how to ape someone's style. There's nothing in the code at all.
Click to expand...

Really yes.

They need all that stuff to learn how to do something, they are not cutting and pasting from images, that is why everything they produce is unique.

This is what you get from one of the most popular current generative AIs when you input "the mona lisa" for the prompt.



As you can see it does not reproduce the Mona Lisa.

Even if you try and push the creation "a painting of the mona lisa that is identical to the real mona lisa" you won't get the original Mona Lisa, because it has not got an image of the Mona Lisa stored in its memory.




ETA: Thought I'd see what is meant to one of the best commercially available AI model can produce:

 
Last edited:
Dr.Sid said:
What model is that ? That's .. pretty bad.
Click to expand...

The first two are Stable Diffusion v2.1 using a short run, the bottom one is SDXL BETA short run.

I think it is useful to use a widely known image to show what it is that these AIs do and don't do, i.e. not copy and paste and composite images, these are unique generated images
 
Darat said:
The first two are Stable Diffusion v2.1 using a short run, the bottom one is SDXL BETA short run.

I think it is useful to use a widely known image to show what it is that these AIs do and don't do, i.e. not copy and paste and composite images, these are unique generated images
Click to expand...

Indeed. It's good example for that.
 
Olmstead said:
Really? Gee, I wonder why they need all that stuff then. I guess it's turned into data magic, and the AI just magically knows how to ape someone's style. There's nothing in the code at all.
Click to expand...
Correct. It happens through some magic that the AI's programmers don't understand. :rolleyes:

Darat said:
If you are using someone's copyrighted work it makes not an iota of a difference in regard to the legality of using that image whether you are using it for commercial or personal use...
Click to expand...
In theory perhaps, but in practice it could make a big difference.

Fair Use or Infringement?
Another aspect... is whether the new work is commercial. “Commercial” does not merely mean that you make money from your work. Generally, works of fine art are not considered commercial even if they sell for hundreds of thousands of dollars. Courts are more likely to consider artwork commercial if it is sold as decoration on merchandise, such as mugs, trivets or t-shirts. In that case it looks more like you are using the artwork to sell consumer merchandise, rather than selling the artwork itself. However, the courts are not consistent in this approach. Some courts have held that sales of fine art prints are commercial. Others have found that sales of merchandise by museum gift shops are not commercial.
Click to expand...
 
Olmstead said:
Really? Gee, I wonder why they need all that stuff then. I guess it's turned into data magic, and the AI just magically knows how to ape someone's style. There's nothing in the code at all.
Click to expand...

On the one hand you claim that the images that human artists study aren't in their brains, on the other you claim that the images AI studies must be in their code. I don't see how you square that circle. If it's a logical necessity, then it's necessary in both cases.

The actual case seems to be similar for both, however: the original art influences the state of both the AI model and the human brain, but not by storing the image.
 
arthwollipot said:
Bad is subjective. I think it came up with a pretty good replica, given the circumstances.
Click to expand...

Compared to what I could do it's fantastic. :)

Dr.Sid said:
No, I mean it's way worse than when I tried. But I used offline install with custom model, also as the output is random, he might have just gotten unlucky.
Click to expand...

They were genuinely the first images, I didn't try a few and pick the worse ones. Last night I played around with the prompt and other parameters and I can't get anything that would pass even on a quick glance.

But this raises another issue, the AI art isn't that good or rather perhaps "consistent" may be a better way of putting it. Folk have heard about how they struggle with fingers and spatial dimensions but they make a lot of other "mistakes". Often an image may seem to be a good one on a brief look but when you start to look you see more and more things that aren't quite right. (Which of course you do with human produced image.)

The huge buzz that Adobe has created with their generative fill in Photoshop beta is starting to be more measured and walked back a bit as folk realise the limitations. Yes it is good at some things but rarely does it produce something that you'd say was "final" you are still going to have to do some work to get a finished image. I'd say its main use right now is in concepts and rough work. It's better than dropping an image you've found on the web as a placeholder or if you want to mock something up for a pitch and not have to spend a week producing a concept.

Certainly at the moment it does not replace either artists or even skilled designers using tools such as photoshop.
 
Darat said:
Certainly at the moment it does not replace either artists or even skilled designers using tools such as photoshop.
Click to expand...

Indeed. The next version will get closer.
Current models sucks at hands. They excel at faces though. It's just a matter of focus. Faces were issue in prev gen models. New methods were developed (typically extra network just for faces), and they were more focused on during training. Hands are next.
Alse AI now follows hardware limitations. You need powerful videocard to run Stable Diffusion at home .. and language models are way larger, and those you can run at home are 2 generations behind.
But that is assured to get better. In 5 years these groundbreaking tools will be kids toys. Images will be perfect and in seconds, and there will be video.
As for language models, hard to say. Their training is much more extensive. They will for sure be way better, question is how much, and in what areas.

That's the issue with AI. It will follow Moore's law, more or less. There is no stopping it. I'm actually excited. It might lead to extinction of humans, but it's going to be fascination time.
 
Kinda seems like the problem of trying to take this from a novelty to commercializing it is that widespread adoption of these AI tools will result in less and less human generated material to be used as training material. If you fire all your artists and use an AI art program, what exactly are you going to feed as reference material? What happens when the AI writing software is pulling reference data that was generated from an AI that was pulling reference data that was generated from an AI and so on and so on...

“Over time, mistakes in generated data compound and ultimately force models that learn from generated data to misperceive reality even further,” wrote one of the paper’s leading authors, Ilia Shumailov, in an email to VentureBeat. “We were surprised to observe how quickly model collapse happens: Models can rapidly forget most of the original data from which they initially learned.”

In other words: as an AI training model is exposed to more AI-generated data, it performs worse over time, producing more errors in the responses and content it generates, and producing far less non-erroneous variety in its responses.
Click to expand...

https://venturebeat.com/ai/the-ai-feedback-loop-researchers-warn-of-model-collapse-as-ai-trains-on-ai-generated-content/


AI training data could become more inbred than a Hapsburg royal and could quickly become quite incomprehensible to humans.

These things are just fancy plagiarism machines that are threatening to replace the actual human beings that are generating the work that these AIs need to plagiarize in order to function.
 
Last edited:
Good image tagging requires to know where the image came from. Sure, it's a challenge, but you can always just use the old training sets with new bigger/better models.
The amount of training images is not a big problem I think. There is enough, even today.
The model understanding the image is where improvements can be made.
For example now the language model used in image generations models is pretty simple. You can describe the subject decently .. but only one subject. You can have 2 persons in the shot, but there is big chance that any description of one will also apply to the other.
Another issue is scope. Models like stable diffusion know most celebrities as subject, and many artists as style .. but it can't tell species of birds for example. It know black bird is black .. but that about it. Clearly it wasn't focus. But sometimes it would be useful if he could do that. And that can be learned by studying photos .. or even AI generated images, if they are anatomically correct.
 
What happened when a teacher set an assignment for their pupils to use ChatGPT and then analyse the results.
I had anticipated that many of the essays would have errors, but I did not expect that all of them would. Many students expressed shock and dismay upon learning the AI could fabricate bogus information, including page numbers for nonexistent books and articles. Some were confused, simultaneously awed and disappointed. Others expressed concern about the way overreliance on such technology could induce laziness or spur disinformation and fake news.
Click to expand...
Twitter thread here - https://twitter.com/cwhowell123/status/1662501821133254656
 
Apparently, finding words that usually occur in the same context isn't quite the same as understanding the meaning of those words:
So you think your newfound ability to prompt ChatGPT for AI-generated recipes could result in a culinary masterpiece?
(...)
All this prompting led to what the group described as “hilariously pitiful results.” With many of the recipes, the chef team at World of Vegan spotted ingredient formulations that “would clash right away and where the mishaps would occur.” The team also felt the recipes were largely “deceptive,” seeming ordinary at first glance but often described as “rich” and “decadent” when they were quite the contrary.
Recipe for Disaster? ChatGPT is Tasked to Create Unique, Tasty Dishes and Fails Miserably (The Spoon, May 11, 2023)
Click to expand...
 
You must log in or register to reply here.

Back
Top Bottom