@@Beanpolr most purposes? Dall e has better inpainting and outpainting, and better making photorealistic stuff plus there is tons of site that offer dalle2 free like playground
Hey Nick, could you do a tutorial or a shorts on how to make a simple diffusion model like stable diffusion some time in the future? By the way, yesterday's stream was awesome 😎
@@NicholasRenotte Hi Nick, can you put in your own images to be rendered for eg fashion and beauty? Also can you buy SD yet like you can mid journey? Thanks so much for the vids!
It's most certainly not free, it's free to play about with at home but practical online applications require the API access which is charged on token usage for online image generation.
Hey everyone I hope all of you are doing well. Can anyone help me with this error : RuntimeError: "LayerNormKernelImpl" not implemented for 'Half' ? I recently download and installed stable diffusion and everytime I click the generate button the above error is displayed
no they used 256 a100 gpus so it took 586hrs per gpu time therefore across the span of 256 gpus they were able to train 150,000 hours of data between them all, so letting them run for 24hrs a day for 24.4 days and this process repeated on all 256 gpus u get 150,000 hours worth of training data
@@wainach9518 that's the weirdest time measurement. When you use that many gpus they are all working at the same time in parallel so why not say 500 or whatnot hours instead of 150k. No one cares about that value
If One gpu did this then 17 years in total also imagining that I am 16 and ot would take a year more for that render to be complete dame I mean that render would be going through all my life lol till the age of 17.Also does math work like this or not so if u had just one more A100 can u theoretically cut 17 years in half like something more managable like 8 and 6 month and what of u have 8 of A100 in that case you cut that 8 and a half years in 4 which is 2 years and 1.5 months which is just wow and with 16 of them you are looking at 1 year and .75 month and with 32 of them it is 6 months and 2 weeks approximately and with 64 of them it will be 3 months and 1 week and with 128 of them it will be 1.5 month and 3 days and with 256 of them it will be summed upto 3 weeks and 2 days and with 512 of them it will be 11 days and with 1024 of them it will be 5 days and with 2048 of them 2 days and with 4096 of them it will be 24 hours and with 8202 of them it will be 12 hours and so on.... so basically if u want to render it in working hours then u would roughly need 8202 of Nvidia A100 or badically be a millionaire.Also dont judge my math it is done mentally and at 3:48 Am so bascially half asleep lol and if u find a correction comment bellow and have a wonderfull day.
Lol that's not how it works. You pay a service provider for the 25 days. For example runpod or puzl, who own a server farm. Still not cheap but much less. To rent one A100 costs ~$1.5 per hour. So around 225K for the 150K hours. But I guess you get a discount for such a big order ;) And now after establishing it, the community offers their hardware for free to train the AI everyday. I guess they will add some proprietary magic in some months to make even better pictures and videos and go close sourced to make a huge load of money with their proprietary solution. It's a very smart way.
They might but they did open source it. If they made a proprietary add on and charged for it. Someone else could make an open source add on that does the same thing. I don't know anything about the development team, but I find monetization of the software unlikely.
Interestingly enough, that number is way off, too. Because you're not factoring in electricity and the cooling system needed for that, nor the people used to train it. Nor the facility cost, etc. So, this is a low ball. Still, probably didn't reach $5 million, when with that factored in
That's actually really cheap, i mean really cheap. I know so many projects that cost way more that never complete or are just put out unfinished just so the xan click the done button.
They have a long way to go. Presently Stable Diffusion is nightmare fuel, and I'm certain it will always be warped and lack soul. Faces are contorted, Eyes are dead, there's often an arm that melts into another person's shoulder, deformed hands and feet with extra legs. Rooms have Stanley Kubric impossibility. I broke my arm on a bathroom floor and I'm fixated on recreating the incident. I can't be the only one who does strange things with it.
Considering this comment was made 3 months ago, i would agree. If it was 3 months ago. But if your statement still hasn't changed, then 1800 images i created and saved in my PC would be VERY against you. Especially with many extensions, updated models, and open pose editor SD has today, it can create things similar or better than MJ. Yes, unlike MJ, It just requires a little bit of skills and patience that's all.
I am wondering if Nvidia can in lets say 5 to 10 years bring the amount of time down from 150 hours of all gpu operationg to something like 10 Ai based gpus being able to do that it would need about 100 to 10x performance but Ai is the next big thing and if Nvidia does this they will br able to create a monopy where amd even would not be able to compete and they could sell the gpu for tonns of money and still dont get thier customers angry cause its better then buying 100+ of them individually and that money can go back in RAD and into thier pockets so they can focus more on this but at the same time I dont want them to do this as this would eventually make them pull out of the gefore rtx lineup and as a passionate gamer more then a 3d modelist I would rather go with what Nvidia is doing right now and not put all thier eggs in one basket that basket being Ai.If you have read this far kudos to you and leave a like to show that you are big brain giga chad comptionate reader.
150.000 hours? There must be a mistake there, since 150.000 hours are more than 17 years and stable diffusion has only been popular in the last few years. The most famous papers are from 2019 and 2020. Or are we taking into account the pre-trained embeddings it uses and stuff like that?
@@Forester_Master That's probably what they meant, but what was said was inexact. Thanks for your clarification! Using 256 GPUs doesn't even seem that much for such a powerful model.
That's exactly why I'm very reluctant to use AI, as it generates a massive carbon footprint 😢 I wanted to use it to create some educational content on ecology, but this really breaks my heart.
Only $600,000 for something this good? Wow. That's impressive. I would have expected it to cost a lot more than that.
for one run. they had 13 runs
@@swyxTV lmao 😮
Well yeah, they obviously didn't have to pay the copyirght for the data they took to train as they just stole it instead. x)
@@Ew-wth 🤡
@@florianschneider3982 🤡
I’ve been doing videos with Stable and in contact with the dev team. They’re truly amazing people! Love your content, always learning from it!
I think that's 150K hours across 256 gpus. That's 586 hours per gpu, which comes out to 24 days of training. That's not bad at all.
Has to be, otherwise it would have taken them about 17 years to train the model.
Bro a year is 8000 hours a week is 168 hours it like how is 150k hours 24 days
@@shimwalug2202150k hours across 256 gpus
@shimwalug2202 because it runs parallel across over 250+ GPUs my guy. Do you understand how software works at all?
@@rl6382software go brrrr
Zuckerberg should have thrown some money on something similar, he spent 10B ie 10,000 Million on metaverse
Rich people problems
Looks like he listened to you.
Not that much TBH. I thought it might be a few millions. DALL-E costs were in the millions if I recall correctly.
yeah pretty impressive it was done in so much less than dalle honestly
im with you guys on this, genuinely thought it would’ve been waaaay more!
@@ceticx well make sense. Compare to dalle stable diffusion isn't that great tbh
@@ForceInEvHorizon It's a lot better for most purposes though. And most importantly, free.
@@Beanpolr most purposes? Dall e has better inpainting and outpainting, and better making photorealistic stuff plus there is tons of site that offer dalle2 free like playground
150k hours is 17 years, did they really start that long ago or is there something I'm missing?
Equivalent hours spread across 256 GPUs. ~585 hours per gpu.
@@NicholasRenotte didn't get you.!
@@abdulnafihkt4245 it means 1 gpu is equivalent to 585 hours so if there is two gpu thats already 1170 hours or 48.75 days
@@ForceInEvHorizon All GPU Used @ One Time. All Hours
I believe it was 150k hours total but all the CPU were operating simultaneously
Your videos are great help, brother
There is a diff b/w gpu hours and hours according to that info it took 24 days ie 586 hours
It's not like those GPU died after that lol 😂
Power usage
they were rented anyway
150k is the total hours, it take a lot less when the data is feeded in chunks in the gpus.
Hey Nick, could you do a tutorial or a shorts on how to make a simple diffusion model like stable diffusion some time in the future?
By the way, yesterday's stream was awesome 😎
glad you liked it!! I’ll see what I can do, haven’t reallly delved into them yet.
@@NicholasRenotte Hi Nick, can you put in your own images to be rendered for eg fashion and beauty? Also can you buy SD yet like you can mid journey? Thanks so much for the vids!
It's most certainly not free, it's free to play about with at home but practical online applications require the API access which is charged on token usage for online image generation.
Cheaper than I expected
Hey everyone I hope all of you are doing well. Can anyone help me with this error : RuntimeError: "LayerNormKernelImpl" not implemented for 'Half' ? I recently download and installed stable diffusion and everytime I click the generate button the above error is displayed
what a year its been
That’s cheap! Gotta factor energy in and stuff, but that’s amazing value.
You are genius . Keep it up man.
Stable Diffusion 3 was trained on 384 H200 GPUs, don't know the gpu hours
but, you can sell those cards after training is finished....
How to access?
That's barely a months worth of training
Totally doable 😅
It's an old video. It's way cheaper now.
17 years of hours, nice lol
Can you give me this pipe line code....
I would have expected a billion for the suspence you created x)
Man got charisma 😩
Wow, so they started training the model 13 years before the company was founded. That's some impressive dedication.
Yea thats what im sayin'
no they used 256 a100 gpus so it took 586hrs per gpu time therefore across the span of 256 gpus they were able to train 150,000 hours of data between them all, so letting them run for 24hrs a day for 24.4 days and this process repeated on all 256 gpus u get 150,000 hours worth of training data
How do i add it into my discord server?
150,000 hours sounds like a stretch. I mean that's literally 17 years.
It is spread across 256 gpus
@@wainach9518 that's the weirdest time measurement. When you use that many gpus they are all working at the same time in parallel so why not say 500 or whatnot hours instead of 150k. No one cares about that value
@@redmist4963 I know ahah. Sounds more impressive.
If One gpu did this then 17 years in total also imagining that I am 16 and ot would take a year more for that render to be complete dame I mean that render would be going through all my life lol till the age of 17.Also does math work like this or not so if u had just one more A100 can u theoretically cut 17 years in half like something more managable like 8 and 6 month and what of u have 8 of A100 in that case you cut that 8 and a half years in 4 which is 2 years and 1.5 months which is just wow and with 16 of them you are looking at 1 year and .75 month and with 32 of them it is 6 months and 2 weeks approximately and with 64 of them it will be 3 months and 1 week and with 128 of them it will be 1.5 month and 3 days and with 256 of them it will be summed upto 3 weeks and 2 days and with 512 of them it will be 11 days and with 1024 of them it will be 5 days and with 2048 of them 2 days and with 4096 of them it will be 24 hours and with 8202 of them it will be 12 hours and so on.... so basically if u want to render it in working hours then u would roughly need 8202 of Nvidia A100 or badically be a millionaire.Also dont judge my math it is done mentally and at 3:48 Am so bascially half asleep lol and if u find a correction comment bellow and have a wonderfull day.
what did they throw out those gpu´s after training ? or is this the energy bill ?
It's the cost of renting A100s. They cost $10-20k each, so it's expensive to rent them. I'm sure the electricity bill was high too
That actually feels fairly cheap. But that's only the training part. The research and development might have cost even more !
It’s own run out of 13
So around 7.8 mil I might be wrong thiu
But yeah hiring people probably is more expensive
Thats a bargain that basically means a single individual with enough capital could fund it
Lol that's not how it works. You pay a service provider for the 25 days. For example runpod or puzl, who own a server farm. Still not cheap but much less. To rent one A100 costs ~$1.5 per hour. So around 225K for the 150K hours. But I guess you get a discount for such a big order ;)
And now after establishing it, the community offers their hardware for free to train the AI everyday. I guess they will add some proprietary magic in some months to make even better pictures and videos and go close sourced to make a huge load of money with their proprietary solution.
It's a very smart way.
They might but they did open source it. If they made a proprietary add on and charged for it. Someone else could make an open source add on that does the same thing. I don't know anything about the development team, but I find monetization of the software unlikely.
Interestingly enough, that number is way off, too. Because you're not factoring in electricity and the cooling system needed for that, nor the people used to train it. Nor the facility cost, etc. So, this is a low ball. Still, probably didn't reach $5 million, when with that factored in
But if you own the 256 A100s (~$2m) and the electricity required, it costs nothing.
Worth it
the electricity cost tho...
That's actually really cheap, i mean really cheap. I know so many projects that cost way more that never complete or are just put out unfinished just so the xan click the done button.
MSFT paid 10B USD
NOT FREE YOU COLAB WITH GOOGLE PAYMENTS
How is it possible it took 150,000 hours to train like a year is only 8000 hours.
wow by your 150000 hours well i guess they started 17 years ago
17 years worth of training???
They have a long way to go. Presently Stable Diffusion is nightmare fuel, and I'm certain it will always be warped and lack soul. Faces are contorted, Eyes are dead, there's often an arm that melts into another person's shoulder, deformed hands and feet with extra legs. Rooms have Stanley Kubric impossibility. I broke my arm on a bathroom floor and I'm fixated on recreating the incident. I can't be the only one who does strange things with it.
Considering this comment was made 3 months ago, i would agree. If it was 3 months ago. But if your statement still hasn't changed, then 1800 images i created and saved in my PC would be VERY against you. Especially with many extensions, updated models, and open pose editor SD has today, it can create things similar or better than MJ. Yes, unlike MJ, It just requires a little bit of skills and patience that's all.
ok wow
I am wondering if Nvidia can in lets say 5 to 10 years bring the amount of time down from 150 hours of all gpu operationg to something like 10 Ai based gpus being able to do that it would need about 100 to 10x performance but Ai is the next big thing and if Nvidia does this they will br able to create a monopy where amd even would not be able to compete and they could sell the gpu for tonns of money and still dont get thier customers angry cause its better then buying 100+ of them individually and that money can go back in RAD and into thier pockets so they can focus more on this but at the same time I dont want them to do this as this would eventually make them pull out of the gefore rtx lineup and as a passionate gamer more then a 3d modelist I would rather go with what Nvidia is doing right now and not put all thier eggs in one basket that basket being Ai.If you have read this far kudos to you and leave a like to show that you are big brain giga chad comptionate reader.
people just trolling or really dumb about hour part?
150.000 hours? There must be a mistake there, since 150.000 hours are more than 17 years and stable diffusion has only been popular in the last few years. The most famous papers are from 2019 and 2020. Or are we taking into account the pre-trained embeddings it uses and stuff like that?
150k hours spread across 256 gpu, which is less than a month. I have no idea why the engineer didn't just say that to avoid confusion.
@@Forester_Master That's probably what they meant, but what was said was inexact. Thanks for your clarification! Using 256 GPUs doesn't even seem that much for such a powerful model.
That's exactly why I'm very reluctant to use AI, as it generates a massive carbon footprint 😢 I wanted to use it to create some educational content on ecology, but this really breaks my heart.
Meeeeeeeeeeeh. Playing with numbers to do hype. thumbs down
150k hours? means 17 years years , stability ai started at 2020 , thats fake then
Nothing compares to impact. Lensa AI makes 560K a day.