New open-source AI video generator is INSANE
Вставка
- Опубліковано 11 жов 2024
- Pyramid Flow is now the BEST open-source AI video generator. Approaching Sora & Kling quality
#ainews #ai #aivideo #soraai
Thanks to uPix for sponsoring this video: Generate AI selfies in just 1 click.
upix.app/
Pyramid Flow: Pyramidal Flow Matching for Efficient Video Generative Modeling
pyramid-flow.g...
github.com/jy0...
Newsletter: aisearch.subst...
Find AI tools & jobs: ai-search.io/
Support: ko-fi.com/aise...
Here's my equipment, in case you're wondering:
Dell Precision 5690: www.dell.com/e...
GPU: Nvidia RTX 5000 Ada nvda.ws/3zfqGqS
Mouse/Keyboard: ALOGIC Echelon bit.ly/alogic-...
Mic: Shure SM7B amzn.to/3DErjt1
Audio interface: Scarlett Solo amzn.to/3qELMeu
Thanks to uPix for sponsoring this video: Generate AI selfies in just 1 click.
upix.app/
No Will Smith eating Spaghetti? Useless
I'm sorry, this is a good result for opensource, but when you compare video with Sora saying "not that much difference...". No. They are lightyears apart.
It's true, for the same prompt Pyramid Flow has worse results than what OpenAI Sora showed.
But is it an apples-to-apples comparison on training compute? Or apples-to-apples on inference compute? What about the training data sets and algorithms?
From the research paper it says Pyramid Flow is is trained in China on 20,700 hours on an Nvidia A100 GPU hours".
By the way, China is supposed to be sanctioned for Nvidia chips including A100 so it's interesting they advertise this.
But compare that to Sora. I saw an estimate that "Sora used between 4,200 and 10,500 NVIDIA H100 AI GPUs for one month, with a single H100 AI GPU capable of generating a one-minute video in about 12 minutes, or around 5 x one-minute videos per hour".
So the H100 is way more powerful than A100. And there's 730 hours in 1 month.
So by that math (with likely several incorrect assumptions), it appears Pyramid Flow has been trained on a very tiny fraction of OpenAI Sora.
Sora doesn't exist! It never came out. The best at this moment is Kling.
@@hdfsgervda in terms of numbers -- maybe. But what about underlining tech? You can invest 100x more time and compute into bad architecture for subpar results.
I agree, in my tests this is pretty much unusable, now we're used to the quality of Minimax and Kling. Paid quality beats free sub-par at the moment (I'm not including you Runway, unusable footage AND extortionate prices is the worst combo)
It's not open-source as it doesn't allow unrestricted commercial use.
Thats only if they catch you….🤫
@@theredknight9314 In that case everything is open-source until they catch you.
@@vytahnot if you have to pay for a license.
@@vytah yep
@@vytah how can they catch you????
Come on, this is not sora level! Sora doesn't have as many morphing issues and it's not as realistic.
SORA doesn't exist Lol
@@beyounickvlog5285 So all the film makers and artists that have been given access are just lying, right?
@@1sava that was a joke btw . And yeah those artists create 100 of generations from the same prompts and cherry picks the best one. You can watch their interview.
@@beyounickvlog5285 Fair enough. Cherry picked or not, they did generate hyper realistic generations. But this model is great for the Open Source industry
@@1savait practically doesn’t exist. None of us have access to it.
The industry has been waiting for an open-source kebab video generator.
That wait is clearly over
we must make the most realistic kebabs
Hollywood level movie with a simple prompt in the next 5 years. Not impossible.
It'll be done before Christmas. Initially AI films will have continuity issues but clips will be compiled together to resemble full length feature film formats by indie devs, probably within this month.
Edit: As @MartinZanichelli mentioned, the audio will be a hurdle but Ik my statement to be true because I am going to do it.
3 years
At least 20 years. But Hollywood will become superfluous.
Ya'll are underestimating AI. Mark my words, it will be no more than 6 months.
Ok, but it will take you a lot of time to elaborate a really good script. Plan the scenes, arrange the footage with the sound. It will take you a lot of time and work, but you can do it yourself alone at home.
26 gig memory. Are you kidding me 😂😂😂
$$$$
Maybe we could see book to video in the next couple of years.
that'd be cool
The potential of AI generated videos is truly remarkable, particularly for architectural visualizations and establishing shots in zones where drone flights are restricted. Keep up the excellent work and enjoy the creative journey!
3:57 Sora still the best. However we do not know how much cherry-picking they have done.
Well, seems like at the end, technology came here to be OpenSource... The sora was left behind.
What is going to be fun is when AI video gets to the level that it can be fed a book and create a movie from it. And that will be here soon.
and i am a proponent of that. TOASTS TO FILMMAKING WITH AI TOMORROW!
The tech behind Pyramid Flow is a major step forward. Imagine the creative potential once the consistency improves. Can’t wait to see where it goes from here.
yes! plus since its open source and tunable, im sure the community will improve this fast, like they did w stable diffusion
"Sora Level Quality"? I think you have to rewatch the old Sora videos again. It's clearly far behind. You even show the Astronaut video and it's so obvious that it's morphing all over the place and getting blurry with strange double lines over time while Sora is super stable and clear ;P But beeing open source is of course super interesting!
For open source is good. But is not competitive right now
Попробовал пару промтов и сделать видео из фото, не впечатлило пока.
Ai search knows its not as good but he has to be subtle and promote this stuff to keep the channel going yall dont get whats really going on here (thats why hes showing them side by side) hes actually showing us how good sora is shhhhh!!
The future AI video tech is clearly targeting the future and upcoming Nvidia 5090 cards at 32 GB. I have a 4090, but it looks like Im going to have to sell it soon and upgrade, wasnt planning on upgrading for many years...
2:04 Has four legs. 💀
😂😂
I do believe it takes sora more compute. The latest models don’t need as much. Sora is probably almost two years old. But when they real ease their latest trained sora it will be top notch. They raised compute power to a mega scale on a vid gen with less training sora 1.0 and you see how it looked.
These footage reminds of the early versions of Dalle!
Its only gonna get better
thanks for sharing!
You lost me at “It’s not much different” [from Sora]. Sadly, if I can’t trust your judgment, I can’t trust your channel.
And that "cat" looked pure nightmare fuel. I have to assume this dude have never been close to a real cat.
@@cajampa Ha! Agreed.
You act like Sora is a piece of crap.
Sora isn't publicly available. It's vaporware and if ever released will be wrapped in layers of OpenAI censorship.
There's also no indication of how cherry-picked the examples Sora examples were (though neither about how cherry-picked Pyramid Flow is).
Also we also don't know how much compute each Sora example takes.
Algorithmically this new approach may be equally strong as Sora, just they might not have the compute to make a bigger model
Then don’t watch this.
Looking forward to GTA VI, but we will probably be able to live in GTA VII and live entire secondary lives.
Well the boys at black Forest Labs will have quieter the bar to reach once they realize their AI Video Tools.
Hopeful alongside all 3 versions of Flux 2.0.
That's quite the one Two Punch.
all of this is just so exciting!
By the end of the year this thing is gonna be crazy
yes!
It's not bad at all, but still rough around the edges/details. The general concept is communicated clearly, it's just the details that need some work.
Very happy about it being open source; Now it's your turn Meta/Llama ;)
I wonder if one of the Apple M3 Max chips with 128GB VRAM would run this?
Free looks always better 😂
yes!
Why the chinese are open sourcing it???
*I wouldn't be surprised if its to collect data since well a war with them is on the horizon*
Because of communism :)
@@TheNjordy They are not as greedy as the Americans and smarter.
@@AnimagicToonsThat's not the case for sure.
Because they are smart and think long term. They are better because they have no mixture, no impureness. 🙋🙋♂🙋♀✋
Hyped! DiTs quantize great, so the FP4 version should fit in 26/4 or about 8GB of VRAM. 😊🎉
can't wait for that!
Not Sora level, but still veeeeerry cool.
TBH most of the non-cherry picked outputs I've seen have some pretty bad decoherence, artifacts, and blending
Hmm open Source oh man it has begun
Why are you showing us something that we can‘t even run locally? What‘s the point?
"When open source catches up"
Waiting for meta video generator to be open source
Doesn't work as advertised. The videos i've generated are not really making sense anyone with a solution ? I'm running it on a 3090
OpenAI with Sora showcase videos: Do you want it? Do you want it? 🤭🤭 * never releases it *
That one chinese: 🗿
lol
Test driving NotebookLM to do your voiceover?
And how many hours of barbeque video would it take to train a model to output barbeque with this much freaking fidelity? I mean I can literally taste the peppers and shicken
5B/flux “1.1” model release date?
Am I tripping? What are these comments lmao. Runway and Minimax have far surpassed Sora for a while now. Minimax, especially with this now IMG to Video tool is by far the best, Sora isn't close. Why do you keep talking about some video generator that still isn't out and there's like like 5-6 different ones released since then?
All it took was someone with brains and another game changing of A.I industry falls in the hands of the people.
can we use it locally?
yes (if u have enough vram)
Yup. I'm installing it with Pinokio > Gepeto right now
yes, nice, but how we test this?
The results are pretty good but they're significant errors. For example, The video of the astronaut,. his eyes are messed up. And The video of the cat waking up demanding breakfast... The cat's mouth is a bit deformed.
thanks for sharing!
And I can't tell that it's a steam train. Looks more like several flat cars followed by a pair of diesels.
I don't think sora ever existed lol now we should compare things to kling no more sora 😅
It would be interesting to translate a short tale into a sequence of clips using this
So, it's not available to try online, right?
they just added a hf space: huggingface.co/spaces/Pyramid-Flow/pyramid-flow
I got 1 free video from it. 3s long
😅 Why say about colab and then delete the comment? Works on colab but image-to-video in just over 40gb so A100 won't do it.
cool. how many vids could you make in colab before the limit is exceeded?
@@theAIsearchthere is no limit.
This is (again) one of the game-changers I've been waiting for. I have 32gigs of RAM, but this kind of install is beyond me, at the moment. I've already fbared my main drive with improper installs so I'm going to spend some time straightening that out and trying to learn a few more things before I dive this deep. Still, this is exciting and I can't wait for the updates. The future looks bright.
Couldn't agree more!
Not just 32GB of system RAM, but 24-40GB of VRAM on your GPU.
@@High-Tech-Geek Thanks. I'll check my card specs. I need to do an audit of my resources. I've gone from knowing nothing to knowing a little, since I started this journey.
you are going to need rtx 5090
@@AutonomousUltraInstinct69 I hope those live up to the hype.
What do I think about this? I think all the good in the world! Long live open source.
will rtx 4060 ti 16 gb be good enough for this?
you think that card is good enough for ai generation and voice changers?
nope, not for now. i also have 16g
this is good for images and voice though.
@@theAIsearch I see. I will need something stronger. can I get it to work despite being slower? or at least can I use image to video generator?
@@Mfrt-e7n they will improve it for lower vram. gotta wait a few days hopefully
@@theAIsearch thank you.
let's hope they'll optimize it enough for 16 gigs at least lol
Sora is pathetic. They really thought they did something in February but showing off their THEN boom🎉 Runway, Kling, Hauilo and Pikalabs AND with Meta gen coming up, they all put Sora in hiding😂
yep
Give OpenAI a break. The censorship and political correctness filters won't code themselves
No. It's not pathetic.
Is it censored in any way? Can I generate hardcore waifus in action?
great minds think alike
RTX 5090 32 GB will run it just fine (you just need to pay 2500+$ for it first ^^)
alright, this is what i'll save up for
Turn captions on?
Thanks for the video, thanks for sharing!
It's nowhere near Sora or Meta's video generator/
So how can I use it exacly?
they just added a hf space. literally just now: huggingface.co/spaces/Pyramid-Flow/pyramid-flow
@@theAIsearchokay but it’s not free ? Hugginface have a limit use
Yaaay
@@Noahperaudon i got 2 videos out of it before my free limit was exceeded
@@theAIsearch Yes but well it’s a shame, isn’t there an alternative to use it otherwise?
8:40.
It's George.
George Bush
should be able to run it locally using runpod, will ty it out now
good luck!
@@theAIsearch works perfectly :))
I have 2x RTX 3090, so I can probably run it on my PC in 768p. But I'm hesitant to install anything approved by the CCP on my PC. Kling as a web service is one thing, but this... I'll just wait for BFL to release their own model.
looks like i needa start stacking GPUs. one is not enough
@@theAIsearch don't forget to get a good power supply (at least 1400w)
i think it's good for landscape videos
yes!
Could you do a video on how to do song cover's etc on mobile android and iOS, because my PC broke (best for free) and you could do it on the go as well outside the house, if you do it thanks :)
From Poland 🇵🇱
Hello
yo!
yoo it's the legendary michael superbacker, didn't expect to see you here.
Replicate gonna be making beaucoup dollars
$$$$
Very similar to LSD visuals :D
my 1 tb is full of ai help me!
Thank you.
You're welcome!
Yo
yo
Can it do nsfw
i'll def test it out 😏
@theAIsearch Definitely interesting
Yay !!!
Really good
All looks fake
Awesome
yes!
Too many ads. Your channel is just not worth it.
:DDD bruh
Bad video quality.
It still looks like hot garbage, not really useful for anything yet unfortunately
First comment
😃😃😃
Second
third
\
😃