Author Topic: Generative AI Collection of links and important articles, videos, court cases (Read 57795 times)

Jo Ann Snover · « **Reply #300 on:** March 07, 2024, 16:34 »

https://petapixel.com/2024/03/07/recreating-iconic-photos-with-ai-image-generators/

None of the copies are great, but they're clearly more than "inspired by" the very famous images they were trained on.

Andrej.S. · « **Reply #301 on:** March 08, 2024, 03:34 »

Quote from: Jo Ann Snover on March 07, 2024, 16:34

https://petapixel.com/2024/03/07/recreating-iconic-photos-with-ai-image-generators/

None of the copies are great, but they're clearly more than "inspired by" the very famous images they were trained on.

Actually they are really bad copies. I don't understand why you would want to recreate old famous photos. It's like recreating Mona Lisa, why should you? What's the point? Famous unique vintage content holds it's value in it's authenticity.

But you can create new original images with new perspectives, new techniques, styles, etc. You would try to evolve and not stagnate.

Andrej.S. · « **Reply #302 on:** March 15, 2024, 07:26 »

Whoa, found today wirestock's AI image generator:

https://wirestock.io/ai-generate

Dam*, that thing is better as I would expect. The generator is obviously using finetuned Stable Diffusion models.
Their price is even not that bad. 10 bucks / month for 1000 generations.

Here are some test generarions:

https://ibb.co/VCX2VPr
https://ibb.co/dtyxc1J
https://ibb.co/s1fK54j
https://ibb.co/Q9VySpR

This development makes me quite nervous.
It's faster than I would expect.

I can currently beat the generated results by only using an AI upscaler (altough I haven't tried out wirestock's upscaler yet).

As an AI prompter I think you should already now focus on learning to generate visual aesthetic and striking images to stand out of the coming storm.

Andrej.S. · « **Reply #303 on:** March 19, 2024, 08:11 »

NVIDIA Launches Blackwell-Powered DGX SuperPOD for Generative AI Supercomputing at Trillion-Parameter Scale

https://nvidianews.nvidia.com/news/nvidia-blackwell-dgx-generative-ai-supercomputing

Not that I understand much about the technical specifications, but what sounds promising is the following sentence

Quote

GB200 Superchips deliver up to a 30x performance increase compared to the NVIDIA H100 Tensor Core GPU for large language model inference workloads.

So Nvidia released the old H100 just last year and now they release a new for AI optimized chip, which delivers up to a 30x speed up.

Imagine how advanced AI will become the next years if Nvidia will hold on the release pace.

Jo Ann Snover · « **Reply #304 on:** April 13, 2024, 12:41 »

Bloomberg article about Adobe Firefly training - "Adobe’s ‘Ethical’ Firefly AI Was Trained on Midjourney Images"

https://finance.yahoo.com/news/adobe-ethical-firefly-ai-trained-123004288.html

PetaPixel via Bloomberg:

https://petapixel.com/2024/04/11/adobe-will-buy-your-videos-for-up-to-7-25-per-minute-to-train-ai-report/

Bloomberg: "Adobe Inc. has begun to procure videos to build its artificial intelligence text-to-video generator, trying to catch up to competitors after OpenAI demonstrated a similar technology."

More interesting, especially in light of the first article is one of the comments (emphasis mine):

"I just realized what this actually means. It's mostly marketing. Clever marketing.
They can't possibly gather enough material in a timeframe that can compete with simply scraping or buying in bulk. While they wait for volunteers to upload content, the other AI companies will have stolen the entire Internet twice over (which is what I think they are also doing, behind the scenes).
This is made to become product differentiation at launch.
And it's impossible to check the numbers (creators, video time etc) because it's confidential.
Brilliant.
PS: they can use the material they gather through this method, it is useful, but it's just too little. This method provides protection in the case of overfitting, or it can be used to produce better results by referencing more aggressively."

synthetick · « **Reply #305 on:** April 13, 2024, 18:30 »

Some useful insight into how much companies are paying to stock agencies for datasets:

https://www.reuters.com/technology/inside-big-techs-underground-race-buy-ai-training-data-2024-04-05/

"Rates vary by buyer and content type, but Braga said companies are generally willing to pay $1 to $2 per image, $2 to $4 per short-form video and $100 to $300 per hour of longer films."

It seems to me that only a very tiny percentage of this is being paid to artists.

Uncle Pete · « **Reply #306 on:** April 14, 2024, 12:37 »

Quote from: synthetick on April 13, 2024, 18:30

It seems to me that only a very tiny percentage of this is being paid to artists.

There's no doubt about that. If Adobe is true to their word, we are getting 30%, minimum for any image used. Make a note, other like SS and IS we'd get 15%? But I don't see Adobe disclosing what we are actually getting paid for data training.

Training AI using other AI images, is a dangerous tactic. Adobe admits that 5% or less, which was carefully inspected, was used. WHY? First off, if it's such a small number, why not find real images to fill the voids? But 1% is still, some percent, so Firefly is not free of scraped data.

We're still watching the courts to see what is decided about out of copyright and fair use. More claims, more trials, somewhere, someone has to make a decision that defines and decides with clear limits and regulation.

Jo Ann Snover · « **Reply #307 on:** April 22, 2024, 08:52 »

I had missed this last month:

https://www.freepikcompany.com/newsroom/freepik-reimagine/

Freepik's AI collection is already much bigger than Adobe's (84.27m versus 60.8m this morning). I don't know how well it works in practice, but the idea is really appealing "... to allow users to upload a picture and instantly receive a unique and exclusive prompt. With this prompt, users can easily interact with the image and make changes." You find something that's close and use AI to get what you need - no need to figure out the basics of a good prompt as that's done for you.