massive point of consideration, this channels talking head is full of gas.

However, their dataset was not seen by anyone. No one will, either.

@@bonecircuit9123 this KZfaqr's ego has gotten too high 😂

Absolutely no matter what the model is trained on.... No model uses parts of other people's images when generating new ones. So any model is 100% safe for commercial use. But firefly is terrible. SDXL is the best model right now and absolutely free besides.

They claim that,but we have no seen the dataset,so it;s entirely possible it's uses copyrighted images,personally I don't really care but they might have a problem in the future.

its crazy how fast thing are progressing that it only took a year for us to get disillusioned with this

What is your name a reference to?

@@pattyprolapse Insightful overview of the latest AI developments! It's fascinating to see the rapid progress in areas like image generation and upscaling. Excited to explore the open-source possibilities mentioned here.

The entire concept of censoring these models is madness. In 10 years we'll look back at such prudish bs as ludicrous and borderline mad. It's like trying to hold back the tide with magic chanting and slogans.

Yes SD3 API is way too censored even for perfectly sfw prompts. This is why I’m hoping they do release the weights soon

Thank you for pointing that out. Everyone is on the AI bandwagon and claiming that things are moving lightning fast. Excuse me. We are seeing proof of concept items beyond LLMs and image generation. We need the next AI product to actually DO SOMETHING and not just promise to do so in the near future!!

The answer is simple...its come as far as it can go and is fizzling out...!!🤣🤣🤣🤣🤣🤣

Big companies are slowing down and are not releasing because society isn't simply ready. OpenAI has a audio model they haven't released since 2023 because it could create deepfakes. Sora is known to the world, but still not open to the public. Udio, the new music AI service, was created by former Google Deepmind employees, who made this type of technology since probably 2023. The AI Apps we have access on the internet are probably a fraction of what is possible with LLMs inside huge corporations like Google or Microsoft. I think the only way forward is smaller companies trying to get ahead of the competition by making products that bigger ones were not ready to release (like Udio), and then bigger companies feels its time to release them to the public. All in all, the AI revolution is being delayed because governments, the media and society in general because they are simply not ready or aren't aware of the life-changing potential.

This is gonna age poorly when GPT-5 is announced. ...and then the regulations will come, and we will have no more Innovation for a year.

You will be seeing more change once Nvidia Blackwell starts rolling out

so not free at all, lol

@@ouroborostechnologies696 ye, unless you're still in school, so that's their target audience with this

if you edit your comment then it removes hearts by the creator...

@@hypercoder-gamingRookie’s mistake

@@hypercoder-gaming bro…..

Made me laugh

😂😂😂

it may be like the 2000 web bubble that bursts fast but is part of everyday life 5-10 yrs later

Yeah but for long prompts only Dalle-3 really works

@chanpasadopolska Depends.........I access Dall-E 3 via Bing Image Creator (BIC) for 98% of my usage. In fact, I've used CoPilot to create shortened versions of prompts which are more than 480 characters (spaces and punctuations included) as this is the BIC limit. There's very good reasons to use BIC vs any other access point for Dall-E.

naw stable diffusion 3 sucks an its super censored

@@vegax5589 I suppose it's all subjective. I've gotten some amazing results with SD3.......better than MJ 6 is quite a number of goals. Besides, once the model is available via forks like AUTOMATIC 1111 and Comfy & other access points, that will be big.

Obviously hallucination is a problem, but nothing has given me the impression that more parameters = more hallucination. Where are you pulling that from?

@@justtiredthings papers I spent months to read… it’s showing how training can affect models, and more training can kill the performances, but also the number of parameters, the layers, de dilation, and so many more parameters .. The transformer architecture by itself has not been improved since the first models and research is not very productive in a good way. What I understand is that to get more sustainable evolution we could train billions of LLM models with slight changes and let it evolve with better benchmarks. But that would require more than six order of magnitude more GPU horsepower than B100 clusters…. When improving something in the models you have to run very expensive training to see how it affect your model evolution… that’s the main issue to go faster. So decades long plateau is coming. The famous AI winter scientists were talking about 2 years ago… That’s not what I was expecting. That’s very disappointing as I KNOW humanity will achieve AGI in the next century with all the benefits (life expectancy, abundance, big scientific discoveries… maybe the true nature of the universe or the answer to Fermi paradox…) Have to live longer to see that

@@notalkguitarampplug-insrev784 from everything I've seen, hallucination seems to be reduced a bit as models gain more world knowledge and parameters. Small parameter models are famously bad with factual accuracy. But I wish you were right, because AGI is a horrifying outcome, and we should be doing everything we can as a species to avoid creating it.

Wait what. GPT 4 level performance with a 70B model and we're not even past Q2 2024

@@markonfilms i strongly encourage you to have fun with Llama 3 70b using groq. It’s faster than any human could be but still less accurate and reasoning is not on par with GPT4 (and even less with current version)….