This Free Image AI Is Gonna Break the Internet
HTML-код
- Опубликовано: 10 сен 2024
- To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/... . You’ll also get 20% off an annual premium subscription.
With Black Forest Labs' Flux.1 Suite being also added to Grok-2 in X/Twitter, I think AI image generation is about to get to REAL mainstream with how insane the quality it creates. Maybe next time I'll talk about how images are combined in video generators to bring the level of AI visual trickery even higher.
This video was sponsored by Brilliant
Special thanks to neggles for helping this video
check out my newsletter:
mail.bycloud.ai
Black Forest Labs
[Website] blackforestlab...
[Flux GitHub] github.com/bla...
Flux.1
[Flux.1 Dev Huggingface] huggingface.co...
[Flux.1 Schnell Huggingface] huggingface.co...
[Flux.1 Pro Fal.ai] fal.ai/models/...
[Flux.1 Pro Replicate] replicate.com/...
Other Sauces
[Latent Diffusion Paper] arxiv.org/abs/...
[Stable Diffusion 3 Paper] arxiv.org/abs/...
[Diffusion Transformer Paper] arxiv.org/abs/...
[ComfyUI] github.com/com...
[SwarmUI] github.com/mcm...
[ControlNET Canny] huggingface.co...
[ControlNET Alphas] huggingface.co...
[Img2Img @NerdyRodent] • Flux.1 IMG2IMG + Using...
[Flux.1 Inpaint] huggingface.co...
[Flux's Arch Diagram] x.com/nrehiew_...
[SimpleTuner] github.com/bgh...
[Flux-realism LoRA] huggingface.co...
[XLab] github.com/XLa...
[IKEA LoRA] x.com/ostrisai...
[Flux's LoRA on Fal] fal.ai/models/...
[Flux's LoRA on Replicate] replicate.com/...
[@mreflow LoRA example] x.com/mreflow/...
[Running 12GB VRAM] www.reddit.com...
[Running 4GB VRAM] www.reddit.com...
This video is supported by the kind Patrons & RUclips Members:
🙏Andrew Lescelius, alex j, Chris LeDoux, Alex Maurice, Miguilim, Deagan, FiFaŁ, Robert Zawiasa, Daddy Wen, Tony Jimenez, Panther Modern, Jake Disco, Demilson Quintao, Penumbraa, Shuhong Chen, Hongbo Men, happi nyuu nyaa, Carol Lo, Mose Sakashita, Miguel, Bandera, Gennaro Schiano, gunwoo, Ravid Freedman, Mert Seftali, Mrityunjay, Richárd Nagyfi, Timo Steiner, Henrik G Sundt, projectAnthony, Brigham Hall, Kyle Hudson, Kalila, Jef Come, Jvari Williams, Tien Tien, BIll Mangrum, owned, Janne Kytölä, SO, Richárd Nagyfi, Hector, Drexon, Claxvii 177th, Inferencer, Michael Brenner, Akkusativ, Oleg Wock, FantomBloth, Thipok Tham, Clayton Ford
[Discord] / discord
[Twitter] / bycloudai
[Patreon] / bycloud
[Music] Massobeats - Glisten
[Profile & Banner Art] / pygm7
[Video Editor] @Askejm
![Breaking Down Meta's Billion Dollar LLM Blueprint [Llama-3.1 Full Breakdown]](http://i.ytimg.com/vi/68_BKkymYs4/mqdefault.jpg)
To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/bycloud . You’ll also get 20% off an annual premium subscription!
Flux.1 is also now available on X with Grok-2, but I think you need to have subscription to use it
Do you still have the prompt for the thumbnail? Did you use those really long natural language GPT prompts?
@@minecraftmaniac2003 generate an image of a natural looking dark hair Asian girl with taking selfie in front of a mirror in a toilet holding an iphone 14 Pro in one hand and holding a paper on the other, looking at the camera with text on the paper saying "I am not real" that looks hand written and not in cursive. With the image quality looking like its taken in snapchat. Make the selfie extremely mundane and realistic with low contrast and saturation. The girl is wearing a sleeveless black flowery dress standing in front of the sink counter.
@@bycloudAI thanks, flux prompting is weird to adjust to after booru tagging
You cannot be tricking us with those Fireship thumbnails lol
Exactly what I was thinking.
he is better than Fireship
IKR?!
@@theo4884 that's interesting.. how so?
mimicry is the highest form of flattery
4:50 For anyone wondering, Flux Dev has a non-commercial license only for the model and derivatives, as expressed: "You may only access, use, Distribute, or creative Derivatives of or the FLUX.1 [dev] Model or Derivatives for Non-Commercial Purposes". However, outputs do not fall under any license or copyright ownership: "We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes)".
Thank you. I was about to post this as well.
All of the outputs are commercial.
Thanks for the clarifications
like that going to stop anyone
2:57 you didn't butcher it. That's how Germans say "schnell"
yeah i was about to comment the same, really quite good pronunciation actually 👏
True, but he did butcher Ilya Sutskever at 0:33
@@snailedlt That's how Germans say "Ilya Sutskever".
@@zimbu_ you change the placement of letters? He said "Skutsever"
This issue quickly
The video is only an hour old and already out of date, lol. Forge Webui (an A1111 fork) has been able to run Flux for a few days now. I'm sure that more option will be quick to follow.
Also worth mentioning is that people are experimenting with ways to reduce the rather hefty hardware requirements. Going from FP16 to NF4 allows the model to run on low end hardware at much better performance.
I mean he did update the end but yeah everybody in community basically have stopped working on SD or others and put all their efforts on fine tuning, hopefully soon we get a version that properly follows artist names maybe Juggernaut Flux.
NF4 only works on RTX cards right?
@@minecraftmaniac2003 Nope.
You can run Fp8 but better is use Q8 or Q4 - nf4 is outdated and giving the worst quality.
@@mirek190 well i have a gtx 1080 8gb and only 16gb ram, so not sure i can even run it at all. Thanks though
@@minecraftmaniac2003you should be able to just fine the ram will be your bottleneck though not vram😂. I am running flux dev in forge on a laptop with rtx 2060 6gb and 40 gb ram . It takes just a bit over 2 minutes to generate a 896X1152 image. I am running the nf4 version didn’t try the q versions yet
fireship style thumbnail
"Code Report" ahh thumbnail 😭😭
you actually pronounced "schnell" perfectly 👌
I'd expect most people to say it like the English "school" or "schism" but he mostly NAILED it.
Easiest test in the world for a native English speaker who has decent hearing
English is part of the Germanic language family yknow? and schnell is the easiest word to pronounce
@@NeatCrown why? sknell doesn't sound right at all. it didn't even occur to me that there would be another way to pronounce it
he starts calling it "snell" throughout the video
I legit thought this was a Fireship video, looking at the thumbnail.
It is nice to see that there are more and more competitions
I keep wonder why anyone with money would invest it this it has zero earning potential, always better off investing in the shovels this is literally a race to the bottom.
@@southcoastinventors6583 you can find a strategy, like they can charge commercial licenses on their better models. or use the publicity and credibility they get from open models to sell api access for customers that don't want to bother with cloud, i'm sure there are other ways
@@southcoastinventors6583 it's simple, the first person to automate AI research will be the ruler of world. We don't need to automate everything, just AI research. Once it is accomplish, all desk jobs will mostly become obsolete.
6:53 ok, that is actually insane. The only super obvious AI thing is text on the lanyard, but even that is only really noticeable if you try to read it. The big finger on the left hand is a bit too big for this perspective, but that's something that only a keen eye would notice.
The pic at 6:43 is also crazy. If I didn't know this was AI - I wouldn't even guess it.
If by big finger you mean thumb, then yeah
@@-SUM1- ye
The mic makes no sense, otherwise I agree.
@@MorgurEdits if you don't know all types of lapel mics, this can easily come off as one. Especially if you don't concentrate on it.
Forge also supports it now. It's mixed bag. Absolutely next level in potential and detail level. But also very inconsistent. The same prompt can give you vastly different outputs across seeds. Anime seeps in all the time. Depth of field seeps in all the time, and there is no negative prompt. Adding "no" to prompt keywords does not work. It's a fight.
But damn sometimes it throws insane stuff, and prompt following is generally 10 times better than SDXL.
Long live animes
I understand why their pro variant is not open, but it hints at what will one day be the state of open source (much like what happened with GPT-3.5 or DALL-E 2).
The original Stability struggled because they had no way to make money so I think this is a good trade-off where we get powerful opensource models and they can charge for the best one. Flux dev is probably the best opensource model (despite there being an even better version) so I'm happy.
@@N8O12 yeah, bummer that we won't see a Juggernaut, PonyXL, etc equivalent because of the license and possibly technical issues. But what we got is already very useful.
it's not open source or open weights because it's under a custom license that restricts commercial use (and probably other things that the open source definition rules out).
edit: nevermind "schnell" is actually open weights, sorry…
0:33 Calling Ilya Sutskever "Skutsever" is criminal for someone so invested in the AI sphere
3:36 they hands
AI sucks with hands, no matter how good the model is
@tsunami870 ai sucks w/ hands that's true but it does matter how good the ai is
@@tsunami870it wont suck maybe next year
People should use the inpaint tool more often.
on the sd subreddit right now they say that training loras for flux is easier than SDXL, so i don't think you're correct.
they also don't need many images.
His information are just outdated
Bro ai moves so fast a speedy reseacher is out of date
Wow, an AI company that is not from Silicon Valley. Let's hope they can survive 😅
With a pyramid as their logo 🤔funded by establishment guys
3:23
That is way too realistic for my liking holy shit
The text on the shirt is literally a bunch of gibberish
i guess famous people are just gonna have to accept the fact that their depiction online will be totally out of their control?
The real issue is that it can "easily" be fine-tuned and used on non-famous people. Way too much revenge porn potential with this technology. I really hope the legislation catches up.
@@JasminUwU I think this might already be covered? It's not terribly dissimilar, for the purposes of that context, from a skilled artist making photo-edit-based revenge porn.
always has been.
@@JasminUwU It's already very illegal. The problem doesn't lie in the regulation, it lies in the enforcement. Just like 99% of the worlds problems.
Famous people can also claim that real photos are AI fakes, further confusing what is real and what is not to the public.
Catfishing was never more easier...
Man, I wish Civitai could have gotten a shout out for our training. We even have it setup so you can generate with the Lora’s you make.
easy to understand yet technically sound video, SUBBED
One thing: A.I Waifu.
that's the most important the rest is irrelevant
Amen
@@Arthur-jg4ji there are only 2 things that mater.
A. Can it generate Waifus.
B. Can it lewd them.
If either one is no it is irrelevant and will die.
Finaly somebody talks about the obviously future
Come on man, SHE is not a thing lol
Kinda crazy that you can make a picture of Kanye West naked, but not Lana Del Rey. We really do live in a society
I've been using FLUX.1 for some time and it has one prominent flaw. You can't properly control what should not be in the image. As there are no Negative prompting. For example - depth of field, that almost impossible to remove. And CFG "hack" makes the model 10 times slower. Hope community will be able to bypass it, but there is a possibility model will die out eventually (like SD 2 did).
Isn't this currently usable on Twitter/X? For premium price? Was expecting that to be covered in this video. Very insightful still!
3:28 I already love this new tech.
5:05 You are switching up the GPL and the Apache license. Apache 2.0 is a permissive license and derivative works can be relicensed.
WebUI Forge (Illasyviel's Auto1111 fork) also supports Flux, even with NF4 quantization. Just don't use it with xFormers because at the moment of this comment, is broken.
nf4 is outdated. Use q4 or q8 instead.
3:04 THEY SAY a Picture is worth a Thousand Words - and this is like the BG I want - oh jeez - but with a beautiful naked back in the bed curled away from what I'm doing. I once had such a studio and office. I didn't love that it was a 1960's original guest-shack - but the stream running past the fixed glass window and the green were as close as I got to Paradise. Throw in a few Redwood Trees and some early morning mist - and - you get the Picture!
6:36 Why is that N in the second line ?????!?!?!?!!??!
I don't know what you are talking abou
t
IT TRIGGERS ME LMFAO
Don't forget Forge which is making huge steps and also supports Flux.
Bro as a German your „schnell“ pronunciation was literally perfect
Ok natzi
schnell 🐌
Knowledge distillation does not affect LORA fine tuning.
I thought this is the fireship channel
2025 looking crazy
I literally thought why Fireship sounds so different
Don't worry guys I'm sure the selfies they used to train this one on were all in a dedicated dataset and were perfectly legal to use for this exact purpose 👍
they dont even give a damn anymore lol
@@golden-gamingmore5931 g o o d
Yeah lmao. I couldn’t care less about their licensing, just like they couldn’t care less about people’s data
@@user-pt1kj5uw3bwhy should people even have a say in this. AI is revolutionizing the economy. Who cares if they’re using private data
@@user-pt1kj5uw3b to be fair, the person who upload their data to a social network is the one who doesnt care about their data. i guarantee they did not train these models on my photos. because they are not online
2:55 I think you pronounced it pretty well actually! ^^
Bar for native English speakers talking in German must be pretty low. Although probably even lower for Brits
i dont get how else you would even pronounce it? it seems like a really obvious word to pronounce if you just speak english
Plenty of Lora out now on CivitAI plus they isn't really any feasible way to enforce non-commercial use when it is trained on publicly trained images it just legal theater use it as much as you want and if you worried just use img to img at very extremely low denoising strength.
so i can run it on 4 gb vram? gtx 1650 here would i be facing problems in the long run?
Lmao it's hilarious watching you ignore the Fireship comments
💥 This free image AI is truly mind-blowing! The potential it has to revolutionize content creation is massive, and I can see why you say it’s going to break the internet. The quality and speed of the images generated are on another level, making it a must-have tool for creators everywhere. It’s incredible how accessible AI technology is becoming, and this tool is definitely a game-changer for anyone looking to elevate their visuals without spending a fortune. Thanks for showcasing this amazing resource-can't wait to start using it and see the impact it will have on the digital landscape! 🚀🎨
AI just gonna go crazy and crazier
your "schnell" sounds fine, albeit not very scary
Shout up natzi
You pronounced „schnell“ completely fine. Sounded like a native.
So basically you insert the glibadoochi into the phrem and stroke the wakjkhelvbaboochi until the riggamaroochi excretes some pappochi which then takes you to the shopping cart where you can pay the flaboochi or you can read a bunch of flaboochi until the image looks like OMG! AI Images are so amazing, but to get to this part, first return to the jibblejoxk version 5, it has to be version 5 because 4 is a little bit lackluster when it comes to generating frapple. You do not want frapple on your caboochi buffers. Then, if all else fails, try resetting the computer. If that fails, sell your house to buy the latest graphics card and pay someone to setup the frimblemachavoochi so that you can regress back to the tauren tea bag willy manilly got a chilly shripapoochi vlabadoochi. Cue the Low Fi music so it sounds like I know what I am talking about and subscribe.
Not making the dev commercial license and celerity names being removed is walking in the wrong direction, sometimes you like to give a celerity name 10% to add a look to your character otherwise it will take a hundred words.
Wow. Incredible how much you fit into 10 minutes. I did ML research in college not that long ago and crazy how much stuff has progressed since then.
0:49 I spy some Evangelion Battle Music 👀
Acatchpe License: can do whatever you want as long as your derivatives use the same license, and you only use it to make funny / pretty cat images with it.
could someone please tell what is lora?
HOLY CRAB!!!
I took the AI course with Bjorn Ommer himself back in my master. He thought it very nicely and not overwhelming us with all difficult mathematics behind those models. I'm proud of myself a little bit more now with 1.7 from him. LOL. But sadly what I had to do in my master thesis wasn't the thing he was doing so I didn't try to persue to work with him.
FLUX shows how quick the AI and machine learning is advancing. It looks more real than reality. I have no doubt at all that there are going to be significant changes in the world and everything is moving so much faster and efficiently than ever already.
Just arrived to your channel for the first time +1
I've just tried it and it's not as good as the other generators at making a girafe on a bmx.
sorry i dont understand your first few sentences about high school.
i always sat alone
Came for Flux, stayed for mewndane! 3:21
Just what the internet needed, more AI generated images...
.... But, why male models...?
None of the models are free. you have to provide a payment form in ALL of them. Why do you say this was free ?
Is there a source for the image generator Elo graph or scores? I wasn't aware such a thing existed.
This man should do stand up 😂on point
Nice, The AI hype not dead might as well invest again in potential future AI giant
how is it possible to have diffusion without a U-net?
Far out! I know Neggles!!! Lovely person!!!
3:22 I’ve tried those promps on grok and it looks nothing like those examples.
Dev is easy to run at 512 even with little Vram. With 16GB Mac Mini Pro it’s relatively fast with 20 steps.
fk the speed of this video. Are you guys in Silicon Valley eating up MDMA like chocolate?
Do you know that the new model, Groq 2, the image generator, is actually using FLUX 1? This is why it's uncensored and can create any images you want.
that's good news! Now with that AI I can create fake images of famous and important people abusing children to publish them on the internet and tarnish their image forever! thank you AI makers!
You said the german word for fast correctly
thanks so much. God bless you
It's not free. They start charging after 40 photos generated
He means it's open source, meaning you can download the model within your own local machine and use it there entirely offline for Zero cost.
If you are using the image generator from a cloud service of course it is going to charge money, otherwise it wouldn't be feasible
8:05 "this guy reverse engineered the architecture for this model just by looking at the source code" hackerman😎
I'm currently using a ton of LoRa and ControlNet with JuggernautXL but I think Flux will be my next model
Wait… this isn’t Fireship
I want the AI waify
You pronunced schnell very good, it passes as a bayerish or swiss accent
flux on a 3060 and ryzen 5600? that's almost exactly my setup!
Couple of corrects here.
The way you explain distilling a model is just one way its done, we have no real way to know how its done for FLUX.
You can do LORAs on flux dev, they are better than any SD lora so far. Take less images, and less compute.
It's a nice news: They apparently held back their best ideas and brought it out only within a group they trusted. Or maybe that is the news we want to hear...
Nice thumbnail
good video mate
You think 10 minutes an image is bad? I'm using a 1070 at 2 hours per image.
It's empowering to hear someone with a speech impediment narrate with such confidence
Sometimes I wonder if bycloud uses a voice model of himself for these videos entirely now
Like, mündane
If you don't have a 3090/4090, give up.
CRAZY!!!
Very nice presented! 😊
where is the girl from the thumbnail?
Our powerful free AI image generator will change how we produce and consume visual material. SmythOS can assist you with utilizing AI's ability to transform your creative process.
WTF, I thought this was FIreship.
I don't think breaking what we've become dependent on is a good thing
I'm more interested in ho it cs do higher level math like graduate level meaning math can b accessible to everyone even obscure topics
not running on my M2 Air ! Lol .I Need a GPU!
Love me some PERSONALIZATIO
N
the internet is already broken
Ycombinator... They're scsmmy
Invent your own thumbnail style and do not steal from Fireship!
nice fireshipish thumbnail.. good video though.