The Current Absurd State of Generating AI Images
ฝัง
- เผยแพร่เมื่อ 15 มิ.ย. 2024
- To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/bycloud . The first 200 of you will get 20% off Brilliant’s annual premium subscription.
[LoRA] github.com/cloneofsimo/lora
[LyCORIS] github.com/KohakuBlueleaf/LyC...
[SDXL 1.0 Blog] stability.ai/blog/stable-diff...
[SDXL 1.0 Base Huggingface] huggingface.co/stabilityai/st...
[SDXL 1.0 Refiner Huggingface] huggingface.co/stabilityai/st...
[ComfyUI] github.com/comfyanonymous/Com...
[Fooocus] github.com/lllyasviel/Fooocus
This video is supported by the kind Patrons & TH-cam Members:
🙏Andrew Lescelius, alex j, Chris LeDoux, Alex Maurice, Miguilim, Deagan, FiFaŁ, Daddy Wen, Tony Jimenez, Panther Modern, Jake Disco, Demilson Quintao, 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
[Discord] / discord
[Twitter] / bycloudai
[Patreon] / bycloud
[Music] massobeats - lucid
[Music] massobeats - gift
[Profile & Banner Art] / pygm7
[Video Editor] @askejm
0:00 Intro
0:42 A Quick Recap
2:09 LyCORIS
4:58 More Sick LoRA Usage
6:29 "Text" to Image?
8:25 SDXL
11:48 The Weeb Segment
12:23 SDXL Bugs & GUIs
13:54 Sponsor - วิทยาศาสตร์และเทคโนโลยี
![Fc ครับ ขอแนะนำให้รู้จักกับ "น้องเปียโน" l [Nickynachat]](http://i.ytimg.com/vi/HQU6Q_IYTLo/mqdefault.jpg)
Oh ya, I'm currently looking for a video editor, don't need to be highly experienced but just need to have some knowledge in AI. Dm me on discord if you're interested!
and check out brilliant.org/bycloud to start your interactive learning journey 😎😎
PS. Next video 100k Q&A (finally). Throw some questions at me down here!
CORRECTION:
12:11 Waifu Diffusion 1.5 isn't the base model for any of the finetune and merge for anime on civitai
WD 1.5 is based on SD 2.1
Most (All) anime SD 1.5 models are based the leaked NovelAI checkpoint
thanks ash
congrats man on hitting the 100k milestones..mind sharing the play button image? 😁
Damn, looks like the old editor is getting kicked out for the callout XD
No problem bycloud :)
Yo. I just added you on discord. If you still need a video editor, let me know!
Q: Text to video timeline and impact
AI Influencers being followed by Bots is the ultimate irony.
Essentially making advertisers money fall into the hands of social media companies without providing any service. 🤔
They don't need us anymore :D@@carlpanzram7081
Making money without doing anything is what capitalism is all about baby! @@carlpanzram7081
Nonsense. The ultimate irony is and remains humans drawing trees on paper.
That and the National Organisation for Women's logo being an actual clothing iron. That one honestly made me fall of my chair the first time I noticed it.
hey man they're just sucking up all the fake printed cash using fake ads and getting traction through fake followers. its beautiful
The "AI girl" that appears so commonly is what gives it away
Wich one?
Whats her @ asking for a friend
@@lukadragar4596ur mom
@@Dindonmaskershe's the same in all styles, somehow, the photorealistic asian girl who has abnormally huge eyes
It's always the same girl, I know which one you're talking about lol 😆.
Even if you generate Cleopatra, even her face becomes similar to that common Asian girl we all know about.
The fact thas all these complex mathematical models and high IQ AI networks are fueled by pure weebery is the most amazing part of all this.
People used to fear the AI singularity would be the product of a new type of "Manhattan Project", revolving around usage as a weapon. Look at where we are now.
Not really. It's all they do all day, besides watch anime 😂 if you spent all day doing something, you'll be good at it.
@@AdrianMark what a bullshit. you never had passion for anything.
@@Tera_GX I'd not to be too concerned about that. If you are interested in why, read a book called The Myth of Artificial Intelligence (why computers can't think the way we do). It presents an insiders view of the real state of AI.
@@AdrianMark If you spend all day watching anime, you'll only be good at watching anime, not creating AI for generating anime.
I remember a line that I once read on the internet that "All technological advancement will always lead to porn."
I beleive that comes from the 47 "rules of the internet" a now famous list I first saw around the year 2000.
#34. There is porn of it, no exceptions
#35. If no porn is found at the moment, it will be made
Cavemen absolutely drew boobs with a stick.
@@thebull3206 I think nearly everyone knows rule 34 by now...
It mostly leads to us slaughtering our own brothers
Imagine inventing cure to cancer and then watching as cancer statistics absolutely don't change but instead somehow porn and military technologies advance yet again
I know what my granma feels now when I teach her about "new" tech of my time. This Ai Witchcraft is beyond my understanding.
One thing that wasn't mentioned is that SDXL's strength isn't just the higher resolution and trainability, but that it can handle more complex concepts better. Like if you just look at portraits of people, it won't seem as impressive, since a lot of the 1.5 models have optimized for that use-case. But if you ask it for something weird like George Costanza as a pokemon wearing a pirate costume, it'll probably do a lot better job than 1.5 would (no idea if this specific example would work well, but you get the idea). There's things that people have relied on LoRAs or in-painting for, that may just work in SDXL. Probably lots of meme images approaching once people realize the kinds of things that can be combined easily.
...or just prompt - a car made of chocolate ... SD just fail , SDXL make it perfect
its still terrible since the text encoder is like 800m parameters while decent llms are around 10x that
Yeah but it just uses so much more storage. If most uses are met with 1.5 then it won't be adopted any time soon.
@@incription Dalle-3 is supposed to be using GPT as CLIP. But there are obviously different aspects to this, because merely being accurately expressive is just one aspect of an image generation model. For example, people are now also getting good at using very specialized SD1.5 models to "refine" standard MJ generations, to amazing results. The idea is basically using MJ's compositional quality and combining it with the detail rendering quality in a finetuned SD model. SDXL basically already does that, and it shows. Also, it is possible to add expressiveness and specificity to SDXL through LoRAs and embeddings, as people have done with 1.5. Comparatively speaking the CLIP in SDXL now is miles beyond what 1.5 used to be, and with StabilityAI getting into LLMs themselves, I have high hopes that the next version of SD will actually feature a much stronger CLIP to compete with Dalle. For the mean time, SDXL is king with its detail generating quality combined with its expressiveness and compositional quality, especially given the availability of LoRAs, NSFW capabilities, controlnet, and other opensource tools like animation extensions and in/outpainting capabilities. The open community surrounding it is what makes it strong, and MJ and Dalle3 will forever live in its shadow.
@@zxbc1👏👏👏 human intelligence underrated. Compute overrated 😥
Everyone is talking about Ai, you actually explain it. thank you.
They didn't explain it, they just explain at a surface level what it's destroying by stealing and generating copies of itself onto itself. AI is a cancer that should have been heavily regulated and controlled before being unleashed. It's an absolute fucking end-game disaster.
i mean he didnt really explain it. i understood quite a bit since im studying this stuff but he barely scratched the surface
You probably want to check on AI images projected onto 3D models, sht is gonna get wild
UV unwrapping nightmares beyond my comprehension? Sign me up!
@@quazar-omega you have the UV first, then project
@@hugoruix_yt995 ohh, that's great actually
Can you share the name for any of the models that do this?
I used this for maya> cgmGrAIBox
@@sirtimatbob
This reminds me of looking back at your old HS or college papers, remembering you wrote and got decent grades when you totally BS'd your way through 'em, but when you read them now you think, "wow, it actually almost sounds like I knew wtf I was talking about," even though there's nothing to be learned there.
This video is that.
as someone who's been out of the AI scene for 6 months, I have to disagree with you. This one short video helped me catch up with what's changed. (A LOT)
Thank you for saying that, at the 6 minute mark I had just exclaimed "this means absolutely nothing to me" :D
@@vytros12The scary thing is that there's so much fucking BS going on out there right now, that people legitimately just can't keep up. Not even the people supposedly engineering these tools know wtf or how TF a lot of them even work anymore. Some of the main AI/ML engineer teams have admitted openly to not knowing how some of their models even arrive at the conclusions that they do. Humans have naively unleashed absolute chaos into the world, and they just keep making it worse. I don't know wtf timeline this is, but we are absolutely fucking screwed. AI/ML has already been consumed by companies looking to absolutely gut the financial sector. Things like real estate are already being ravaged by AI literally telling realtors and landlords to push for as high of prices as possible (and they're listening and going YEA YEA MONEY MONEY). Just wait until hackers and cyberwarfare organizations really get going on AI infiltration. People are absolutely fucked, and we're all going to suffer because of some greedy morons pushing all this shit. There's a reason the 'inventor' of AI quit and said you're all insane earlier this year.
@@vytros12 right, I thought AI art was gonna make it easier to just get a nice picture quickly but it seems immensely complicated.. all of this info makes me feel like every random internet user must have an IQ of 2000 and I'm stuck here not understanding anything
This has been my experience. I got interested in it for illustration and game dev reasons, but its such a pain in the ass to get something good, predictable, and that also doesn't look like all the other AI art that can immediately be IDd as AI. Now that its everywhere, the internet has been flooded with these types of images and its trendy and I don't like following trends and I don't want my art to look like everyone else's.
I decided that wasn't worth it, and instead have been teaching myself 3d modeling and digital illustration which has been very difficult but FAR more rewarding than the tedium and disappointment of prompting plus I don't really like the fact that so many people aren't getting compensated for having their work trained into it. I also worry about 1) backlash from having used assets that were created using AI 2) getting into trouble down the line where copyright laws are updated and then all the assets I generated are no longer legal to use in a commercial product. Nah, I think I'd rather just learn a new skill which is far more rewarding in the end. This doesn't actually end up saving me that much time, unlike 3d animation which solves a lot of the problems around 2d cel animation and instead just replaces one headache for another.@@slaveNo-4028
I just hate how AI images are being used for propaganda, and how social media and art websites are getting flooded with AI images
Trump-fans can finally show their hero like a muscular superhero of their dreams, as opposed to the fat slob he's really like.
"Art websites" hahahahaha.
That's inevitable with any new technology. Somebody is going to use it for evil. But we shouldn't let that stop us who use it for good and enjoyment.
At this point, I am not even pretending to know what is going on here. Just give me pretty pictures.
fr
I am equaly glad and disapointed that this technology is mainly driven by producing images of beautifull women
Ai will make the porn industry billions of dollars in the next couple of years 😂😂😂
Have you seen Facebook ads? AI-gen'd BS is already making them billions.
@@garygarcia3142
Never in my life have I felt like a boomer as much as this video has made me feel. My brain cannot comprehend this information. Its like listening to a language I don't speak, but backwards. Do I understanding anything? What is technology? Truly a humbling moment. XD
I'm a university student and had my first introductory courses into implementing AI earlier this year. The thing you need to understand is how an individual neuron works in a neural network. Everything they do here is just repetition of that on a massive scale (say hundreds of millions or low billions of neurons).
AI "Art" has no cretivity...those are all copypasta composites of peoples photos and true art.
Man, thanks for these videos.
It's REALLY hard to keep up with so much stuff going on, but AI still feels frightening lol.
also, props to the editor, nice job man
This is terrifying for artists, this isnt a tool to help artists, its a replacement
NOTHING TO FEAR BUT FEAR ITSELF. What do we all fear ? the unknown. We deal with it in many negative ways, like believing non truths, or disbelieving something right in front of us. Subconscious self programming basically.. copium. Some data we just don't understand. Ever hear about the first ships that came over the horizon and the natives couldn't see them ? the visual data was something completely outside of their daily reality. In smaller ways this happens all the time. Anyway.. finishing up my essay reply back soon
Web3 fixes this. Decentralized identifiers (DID) and watermarking for artists to sign their artwork with their private key. The current state of identity solutions is very user UNFRIENDLY because you have to protect a set of words or a long string of hex chars. Far too easy to lose, or be extorted/kidnapped etc.
After a lot of research and academic verification, patents for new hardware were filed, and it has been 7 years of waiting.. the hardware is not finished yet. 2024 timeline on track still, event devices might already be in use, physical high security locations will be all using it sooner or later, no doubt. The first ever tamperproof identification with attack countermeasures, and ZERO DATA storage. Quantum safe too, every transaction is it's own block and uniquely encrypted, noone can access it without "signing" the transaction with their finger. If you are stressed, or have no pulse, the scan will not allow you to get robbed! social engineering is all we will have to deal with. Things are getting complex, people need a new internet safe by default.. and human verification for eg TWITTER is long overdue.. verified accounts doing scams.. charging for a checkmark is just a band aid, and now can give a false sense of security !
Botnets are the biggest problem with social engineering attacks, creating countless new accounts on socials for years and trying to look like real users. MANY have convincing looking history in isolation, only by association can they be confirmed as bots, eg participating in one of those long back and forth "financial advice" bs comment threads, tells you an account is owned by the scammers.
Archethic Foundation in Switzerland.. we aren't far off getting this security. DID identity ? AE will let us have eg our BITCOIN private key, as a child in a DID tree. Your finger will control access over BTC as well, and value limits for whether a finger scan is required is also easy. Smarter Contracts they call it. Creating new DID identifiers as child pseudonymous identities for various sites is possible, but because those are all clearly ONE human the chain can automatically give a website the option "1 human 1 account" or "no interactions between identities with the same human parent... but the website ? static websites can be deployed on chain, testnet has that already. AEmail coming next.. that could be handy huh ? zero spam by default !
Wouldn't it be great if every AI would get their own DID. It isn't finger generated, so it can't be a human... anyone who wants to pretend.. has to be the parent of the DID tree. So eventually we could just put a scanner in every court room (any device works for any person). The deterrent factor would quickly become self fulfilling. No-one can ever hide if they are summoned to testify ! it would be as simple as.. raise your hand.. ok never mind just scan your finger for the jury.
Computer is this the human biology parent of the accounts in question. Y/N.
See that on the news one day.. you would take notice ! I expect it will happen one day, people always push their luck and think they are smarter than the defenders.. organized crime has had well over 20 years leveraging the crap internet, finally defenders will have the upper hand. The only thing that will limit adoption is fear. I would post the reveal video and the 2017 Viva Tech conference interview but this is already plenty long enough and might get deleted by the TH-cam panic basket that ignores all the actual scammers..
@@goldnutter412 quit the zaza
You're saying we shouldn't be afraid of AI... by referencing the indigenous people of America, who were mostly wiped out by invaders??
You're accidentally correct; the technological progress will happen, but it won't benefit any current artists. It'll benefit the people who are trying to eliminate artists.
"But AI can be used to help artists" Sure, and the colonists could've helped the natives develop technology themselves. But the colonists were there colonize, just like tech companies invested all that money so they could reap the rewards
I just want to appreciate what you're doing. Your content is one of the best out there that covers AI and it's research.
Thank you!
Unless you are on computer 24/7 learning and using this stuff there's really no point to in your spare time since things will change and get updated by the time you think you know what youre doing
For me the sign of a good neural generator is the ability to generate "a city on wheels" from a prompt. Everything I've seen so far only gives me bikes and buses.
how'd you decide on that prompt?
Most likely you only tried models in free acess.
Midjourney and especially DALLE-3 are still not perfect, but scary good at understanding what do you want from them.
Maybe you didn't use the right prompt - simply explain in details what do you expect... It can't read your mind (yet)...
@@AstralTraveler "simply explain in details what do you expect"
Pretty sure he expected "city on wheels".
Just google Ron Herron. Bro got you covered with his citys and yes, slave trade is also implied.
The pause for "Lycoris" got me ~ hook line and sinker.
Here´s my take on "AI" as a 3D artist. What it is good for: automatizing boring tasks like UV unwrapping, making textures in different resolutions, sorting complex node systems and trees within a project.
What it isn´t good for: Everything else, especially anything that´s worth to be sold.
I'd argue it's good for two other things: 1) automated brainstorming for idea specifics, and filling in for beginner artist skill deficiencies for those of us who are learning to draw digitally, but still prefer to produce good images instead of crap during the process. I've found much, much more motivation to continue practicing when suddenly 6 months ago I discovered that I was able to skip my utter crap drawing stage and see what my scribbles could be with some skill. And even now that I can draw much more on my own, the AI can still help generate ideas when I am considering composition directions and details. I'm sure to an experienced artist these things seem like crutches, but I would respond that crutches have their uses, too.
Oh, and upscaling old terrible-quality images, though that's more of a utility. That said, I totally agree nothing should go from AI to sold directly.
@@admthrawnuru Good points. I skipped the whole drawing one and iterating for me is quicker if I do it in my head, though that´s probably different for people without ADHD as well, so I might be a bit biased here.
I usually work with reference from other people in the client´s team or sometimes ask my gf if I´m feeling like getting stuck.
Probably doesn´t matter much anyway when companies start incorporating AI as a standard in the workflow.
These are both excellent points and it sounds like you two are using it as a tool (and inspiration) like it probably should be. Where I get concerned is that it's essentially an evermore sophisticated content imitator that, by its nature, has to steal others' inspiration to exist.
And it's big $$ b/c it can also be a tool for displacing the most creative people.
@@NotSoMuchFrankly Absolutely valid concerns, though there´s other possible outcomes, too.
You technically _can_ ethically train AI with your own portfolio or free data, depending on the context. You don´t need 10 Million pictures of big booby anime girls from Art Station, to train an AI that generates realistic tree textures.
And it will not replace real creativity either, because creativity isn´t just kitbashing random ideas together. It requires fundamental understanding of art, ideas and real world context, which aren´t the strengths of soulless machines.
You are spot on about corporations *trying* to replace/reduce artist in their workflow though, but I personally hope that will only lead to creatives fleeing the industry and flooding the indie market.
I see 2 potential problems where it can really go south for us from here, one being Governments not able to regulate AI into the existing law framework. The other is consumers just accepting crap quality so copy paste products become the new standard. Let´s call it the "Marvel Syndrom."
Otherwise the majority of creatives should be fine, especially when they do not let themselves be pushed around by the big corpos.
thank you fr. ai was used before this horrendous trend started and never used other people's work without their permission
Amazing explanations, things are evolving so fast it's pretty hard to keep up with all the new techniques ! Thanks a lot :)
Do not worry. You will be assimilated or recycled very efficiently.
@@Dowlphin Pipe down with your edgy BS, mlp pfp.
@@olivercharles2930 Its not edgy anything, its whats happening in real time. All you need to understand, is that this system is based completely on profit. Thats the only info you need, to understand whats happening and whats going to happen. I hope AI takes everyones jobs, So people can finally stand up and rebel this trash system we live in. Humans are meant for so much more, then being glorified slaves.
Great video! Loved the explanation, keep it up!
Way over my head…but the part I understand, I liked. Thanks for your hard work in creating this video and explaining what is going on.
AI art is an affront to the artists' trade.
I tried a bit with SDXL, and even without all the extra specifically trained models, it did a better job at generating what I wanted (note that I am quite the noob and dont know how to use all the extra tools 1.5 has apart from a lora or 2)
Just more study 😊
12:11 dunno why this caught me so off-guard, but the thick edge lines threw me for a loop for a solid minute. You can see subtle differences in line width throughout that can give it away as AI, but to me this was a big jump from photorealistic and vaporwave to a more cartoon-y style
We are absolutely fucking screwed, and this video is one of many that shows and explains why.
the age of machines has begun
It's interesting how much work we are willing to do in order to avoid doing the other form of work.
Technology is always like that lol
Loser always like that
Is it really work if you are doing something you enjoy? I buy and sell motorcycles as a hobby. I may spend an entire day going and getting and bringing home a bike. Some would say thats a lot of work. I don't consider it work. I consider it entertainment and adventure. I will then do work on it such ad changing body fairings, changing the oil, tires, etc. Again for some thats lot of work. I don't consider it work. I consider it putting together a puzzle. Playing with these AI generators and seeing what you can get out of them for many is not work but entertainment. Its not the end result but the ritual of process and learning that is the reward.
@@spartanx169x Preach!
@@lolvanced Do you buy your own groceries? Personally, I forage for my food like a chad, anything else is the loser way
I’m honestly amazed and scared at the same time. Not only is this going to heavily impact creative’s jobs but I imagine influencers will be at a loss as well (why sponsor anyone when you can generate one in VR?), con-artists will use deep-fakes to scam people online, or people could generate entirely fake events that never happened to slander someone or worse.
I don’t believe there’s any regulations in place to prevent these scenarios, the future media might be a real clusterfuck.
Yes, you are applying critical thinking. AI is a net negative for the world, but most people are so enamored with it they can't project out a few years.
AI will affect everyone,
but because most people are not affected yet,
therefore people will kept wishing people that were affected to be screwed by AI.
until before they know it, they finally become the victim as well.
what those people didn't know is that AI doesn't need them,
AI already collected ton of Prompts from people
to the point that they could just generate their own prompt that was
way superior than yours.
that's why those people are so generous, giving everyone free AI generator, to farm your Prompts,
so that they could disrupt you and makes AI function without you in the future
Free AI generator was used as a bribe, so that they thought that they're the special one,
they thought that the ability to write a prompt makes them better and more open minded
than people with university degree or phd.
when in reality, everyone can write a prompt, even a baby can write one.
therefore, everyone will get disrupted in the future, it's just a matter of when.
remember that AI didn't need human to function.
creatives and influencers do not matter
@@Magicwillnz I wish I could go back in time and stop the automotive revolution... Just imagine all those people affected by it... Losing their jobs... Horse trainers, horse sellers, carriage builders, horse saddle manufacturers... Truthly, combustion engines and automotives were a net negative for the world
Creativity would find a place. And the fakes and lies will eventually be deal with somehow. At the end of the AI is confined to the digital space... at least for now. Giving normal people the ability to make "art" by themselves is actually good, more freedom to everyone. Artists would probably adapt somehow, art can go beyond the "beautiful thing jaja".
Nice roundup!! 🎉🎉🎉
This shit is getting out of control
Yes it does.. and still not enough people are taking the huge problems it will cause in the future serious enough.
Really love this ai news format which goes describing what's new with technical details
"Ai can generate things your brain isn't capable of imagining" what a gross lie
Blud u are thinking that you can outpace a GPU on your dreams lmao
@@flynntaggart7216 Yes, easily
well AI art excells in some topics and the amount of detail is often impressive. I wonder how's that even possible
Pace has nothing to do with imagination. And imagination isn't just mean complete random and chaos.@@flynntaggart7216
Brain is a pretty weak tool, you gotta accept that sooner or later.
6:44 Any link to the source of the high-res image shown here? I couldn't find it via image search
That was the most accurate pronunciation of the word "focus"😁
I'm just glad some fellow weebs are doing the hard work for some one like me that understands 10% of this videos...
They aren't doing actual research though for the most part, just fine tuning models for their specific needs.
@@games528 So... they are doing actual research.
@@olivercharles2930 It's not research to download a bunch of pngs and click on a button that trains an AI model of them.
@@games528 Which is, doing, actual, research
@@callmeandoru2627 They're not the ones with a PhD researching and writing papers on generative AI models. They're spending 10 hours writing a prompt to create sexy anime girl.
Just remember: anyone who is making money from AI generated images - you can take those exact same images and make money off of them.
This isn't the future I want...
@@joelrobinson5457 As long as the technology exists it's better than the alternative.
@@futurestoryteller what's the alternative?
@@usernameluis305 Is this confusing? A world where people with no talent succeed through sheer ubiquity, flooding the market with low effort sludge powered by autonomous "content generators" and the only people who really make money are greedy tech bros, who've proven they literally don't know art from NFTS
Clear enough?
Another way to put it is we _currently_ live in a world where your artwork is theoretically protected... as long as you are an _actual_ artist.
@@futurestoryteller You are the reason I don't like most of humanity. You're that guy. The "I will squeeze myself into this tiny bottle and I will force myself to like it, because that's how things are now."
Hey bycloudAI, i think i have seen your videos about the LoRA tutorial, and i cannot find it now. Is there any way i can access it now? Maybe from patreon? Thank you @bycloudAI
More 2D waifus. I see this as a complete win 🫡
touch grass nerd
@@krishp1104 k, now what?
@@ginqus I take a blade of grass and EAT IT!
@@ginqus now talk to a women
a "women" @@krishp1104
9:14, comfyui generates at way faster speeds than a1111 even for low vram users like me. an sdxl model can generate (at best @40 steps) ~2m instead of the 15-1hr
yea comfy is so much better
Bro, half of my TH-cam ads are these degenerate softwares 😭😭😭
5:10 what platform is this? I don't recognize the UI :D
This cleared up so much for me! I had all these programs and had no idea what any of them did!
Anime helped push AI image generation to the next level. Admittedly this is a lot for me to take in. I think I might have to study AI and pick a specific field within AI.
lol i am doing a degree in ai
@@cpp705 what field are you specialising in?
waifu gen. @@TheAkdzyn
It also helped push voice cloning. All the SVC tech used for AI singing covers was created by Chinese weebs
@@KingMertel this made me happy 😂
2:07 what is that img can anyone send the download link
i kinda wanteed it too :(
I tried to update my stable diffusion but it doesnt work. Now im stuck with the old version. no lora.
Is there a starting point resource available for me to learn how to do all this high-quality image generation?
AI is a solution for a problem we have not found. And it is a solution for a business model we have not figured out.
thanks for this
as an artist it really helps now that I know more about how different generative methods work
still all just a tool and no soul tho
It doesn’t need to have sOuL. does brush have soul? No
@@TheShinorochiThe artist using the brush does though 🤦♀️
@@poodled7794 the same
The immortal soul does not exist in ai
@@poodled7794 Sorry, but still wrong. There is no such thing as "soul".
Do you have a video going over modern AI concepts and strategies in general? I like how you explain things
Porn in 2030 is going to be crazy.
Now I am depressed :(.
Companies, scammers as well as cheap skates are going to abuse it. I wish governments tried to implement rules for these ai models.
You should be depressed that you make such shitty arguments lol
AI is not abusing anything besides people's tendency to be paranoid about the future.
The two prompts that rule them all. Negative prompt Deviantart and Anime.
Makes all images better, even anime.
2:56 what's the website used to filter checkpoint/model used shown in the video?
Yo, I've gotten SDXL to work on 8GB 2070S without as much lag as the video says. I even got the Openpose based ControlNet for it to work with sequential offloading. Let me know if anyone wants a tutorial/repo for it!
Hey if you make a video of it, Please put the link here.
I have 16Gb Ram and 128 Mb vRAM
@@Jadon_WYSE128 MB VRAM? What year are you from?
@@romantroman6270 probably 2015 integrated graphics
Then you're stuck with using the CPU to process.
And it sounds like that would take forever too.
You're pretty much stuck with using the cloud until you upgrade your PC.
@@Jadon_WYSE
Now all those instagram models can go be nuclear physicists and coders :3
sorry if I sound super stupid but is a LyCORIS model better than a LORA one?
9:21 nah mate, that's still many times faster than me painstakingly painting it
DALLE 3 is a game changer. Opensource projects need to catch up.
Well...Dune warned us....
Amazing video as always, could you make a video of how to effectively dreambooth with sdxl?
I did not think I would be getting an anime recommendation out of an AI tech video, but I'm all for it.
The reason why is simple. its because current state of "diffusion" based Ai imagery is 100% limited in poses, and convincing realism with those poses when u use really gimmicky workarounds like controlnet to try to fix the issues with certain poses. The example you showed was using 3D vfx, which creates a much more convincing effect. Ai is just not there yet. Diffusion method is not it.
people want beauty so bad that is pushing AI image to the next frontier.
No it has to always do with speed of fashion. Beauty standards change from generation to generation. Example. Compare beauty from the 80's with 2023, or the 1920's. As we got better tech, we ended up with faster fashion trends that quite frankly don't mean much. It's like movies. Back then, it took time to get a feature movie, today we have a gluttony and nothing is special.
@@SSingh-nr8qz ok
I give you my royalty like, dear sir, for LyCORIS joke 👏👏👏👏
Does anyone know how to generate ai images of a person without the blur background
I've actually gone back to 1.5 models for most of my projects. As of yet I've found SDXL models are significantly better at following specific instructions, but have a severe shortage of creativity. In the sense that any given prompt tends to have a depressingly narrow range of outputs. But even with different prompts there's a strong tendency for certain image elements to look a very specific way way across all kinds of images unless something else is specified. Every model I've tried felt severely over trained.
There's also that these SDXL models really don't like adding dissonant properties to image elements, unlike 1.5 models which can do it surprisingly well as long as you find the right prompt weight balance. Like, if I ask for a cute but scary monster girl, most 1.5 weeb models will find a way to make that work, but equivalent SDXL models tend to just output a cute girl with a scary monster in the background. They don't even try to square that circle.
Also, I've just gotten really tired of the SDXL look. I don't know how to describe it, but there's something about typical SDXL outputs that really started to gets on my nerves, even more so than typical 1.5 coherence problems. Maybe SDXL just requires a completely different prompting strategy, but I've never found one that solves these problems. Well, maybe wildcards can alleviate the lack of creativity. But those are annoying to set up.
What I fear is AI generated images of events that never happened.
Photographs used to be almost enough evidence to prove something actually happened or was real.
Eventually AI generated images can ruin somebody’s life by creating footage of them doing something that they never did.
Welcome to the desert of the real.
u could also achieve that with CGI before current AI era too. ai is just another way to do it
While this may muddy the waters for sure, it's no different than photoshop which can be extremely convincing as well as CGI and visual FX. That's why we have the concept of corroboration, which is the gold standard in the court system (though imperfect, it's still valuable). No what worries me is people willfully refusing to accept corroboration of facts, empirical evidence, and eyewitness testimony and replace reality with their own version of it. This is already starting to fracture society in irreparable ways, just look at the whole cult of Trump and their refusal to accept evidence and facts such as the very clear winner of the 2020 election which was litigated endlessly in court and shot down by a number of judges who Trump himself appointed. Society will eventually bifurcate between those who accept the concept of factual reality, and those who don't. I'm not just talking about people who disagree on how to move forward on sociopolitical and economic issues, but people who disagree if the earth is flat or spherical. Scary times indeed.
@@darkwraithcovenantindustries Who was the legitimate winner of the 2016 US presidential election?
Donald Trump, why? He won by razor thin margins in a number of swing districts in traditionally Democrat held firewall states, supported by people who wanted to try out something different and whose populist economic message was heard and felt by. Because of the nature of the Electoral College, which was the result of a compromise between slave owning states and northern states when the Constitution was first ratified, it is possible for a candidate for US President to win an election without capturing the popular vote, something that has happened 2 times since 2000, and 5 times in US history. Why the fuck are you quizzing me on civics lmao?@@bpj1805
8:11 not even a seasoned artist could figgure out what the hell those promps means.
Would you mind sharing the image from 6:46? It looks amazing :)
All this stuff is honestly so overwhelming that I wouldn't even know where to start.
I'm just hoping that some years down the line we can have easy to use built-in integration of the best AI tools coming out of this into my favorite existing media creation applications (Krita, GIMP, Blender, etc.) to use them to do actual stuff, just generating random images is cool for the novelty factor, but that's where it ends for me, instead being able to have smarter tools to do what I already do, but faster and better, would be an actual boost I'm really waiting to see.
By the way, I love how clearly you cover all these news! It seems so complex I don't even dare keep up with it myself, so thank you!
Depending on how much cost is a factor, you could mess with the beta of Photoshop that includes AI features. I haven't used it yet, but it looks very straight-forward and also easy to put into workflows (since it is PS).
@@ShawnFumoyep adobe dragonfly (the ai tool in photoshop beta) is very easy to use and extremely powerfull and intelligent.
Just selecting something you want to remove and generating without a prompt will most of the time do exactly what you had in mind or better so that's a cool change in normal workflow.
@@ShawnFumo for me it's both cost, unsupported OS and ideology, so no thanks.
Besides, Stable Diffusion is open source, why would I want a crippled version that was tailored to Adobe's taste and politics?
That's why I'm letting it ride and just keeping an eye on it via channels like these. Eventually it'll get to the point where capabilities 100x better than what was shown in this video are just baked in to an ez to use app/platform that I don't gotta spend a weekend troubleshooting just to set it up and a month to figure out how to actually use.
@@ShawnFumo I am gonna guess that since its Adobe, the AI they use will be censored to shit to avoid copyright and other problems, so it would be a very bad idea. Plus Adobe is a horrible, atrocious company, do not give them money unless you're positively sure there are no other options or alternatives.
AI= maths(tensors,matrixs,some pytorch,cuda algortihms...)+art(images,videos,audio...)+hardware(GPu,AI trainers...)
Human= maths(all kinds of neural bullshit) + art (neural bs + external bs) 0 hardware (the brain and all that shit idk)
i dont get anything. can someone point me to a source where i can get started with this ?
All this tech but I still haven't seen a tool that can upscale a blurry video with encoding artifacts and add missing details.
Honestly SDXL creates some really beautiful images. I was on the whole "I think I'll stick with 1.5" bandwagon but after using SDXL 1.0 for the past couple of weeks I don't see myself going back to SD 1.5. Yes images may generate a little faster with 1.5, but I like the images I get with SDXL more. So to me, waiting a little longer for the image to generate is worth it. The key is, IMO, use a model that doesn't require the refiner. There are some great SDXL models out now that don't need the refiner and they render images much faster without it. Two of my current favorites are CrystalClearXL and NightVisionXL.
I mean, we would need another Novel AI leak to make SDXL better for anime, so most people would just stick with older models.
Informative video! Question - can a 1.5 LoRA be used with SDXL checkpoint?
I don't think it can. SDXL is very different from SD 1.5 so a LoRA trained for 1.5 would not work. Also sometimes LoRAs created for different checkpoints of SD 1.5 doesn't work very well on every checkpoint. LoRAs usually work best for the checkpoints they were created for.
No. There is no compatibility between 1.5 and SDXL. However, the most human intensive part of lora training is putting together the training data and getting it all tagged up properly, so in theory, the creator of the lora could take their training data, make some settings adjustment for SDXL training, and then cook up a new SDXL lora. In practice though, most won't bother, and you'll want higher quality training data for SDXL to match the higher resolution, so there would likely be some work going back and reworking the training images (usually cropped and downsized from larger images anyway). Also, SDXL doesn't yet have high quality model variants to train on like SD 1.5. They're getting there (like the mentioned Waifu Diffusion for SDXL), but most early loras being trained now will have bad quality compared to what comes later. With SD 1.5 there was a long period of model variants battling it out for the king of various spaces (mainly realistic, illustration, and anime). Once some high quality leaders were established, lora training had a better base to work off of. So yeah, you're a spot right now where SDXL can't use most loras (since they're SD 1.5) and the loras that it's getting are lower quality than what they could be, and most often, lower quality than SD 1.5's loras. Civitai is pushing SDXL hard with contests giving away GPUs and such, but SD 1.5 is still dominating. Give it a year I guess. SDXL has some big technical advantages, but a deficit of training. A year from now, 12gb and 16gb GPUs should be more prolific and SDXL will have much better tools and model variants to work with.
@@archgaden Thabks so much for your reply mate. SDXL is understandably slower than 1.5 but so much better and easier to use. Speed is a bottleneck. I have a 3090 and I’m complaining. 😅
Sounds like there is a great opportunity for whoever creates a process that can convert SD1.5 loras to SDXL.
@@archgaden
Would you begin learning with SD1.5 or SDXL?
@@archgaden
Editor: So, bycloud, how many memes do you want to use to explain AI?
Bycloud: Yes.
Never forget that the point of AI image generation isn't to "democratize art" but to automate art so that techbros can generate ad content without having to pay artists.
To some extent I get what you're saying but on the other hand I see so many "how to make a game/website/App without coding experience" but never a "How to make a game/website for programming/maths people with no visual design sense or artistic talent whatsoever" and not being able to actually make much use of my years of math and coding practice outside of a terminal due to having such a low artistic/design ability is somewhat demoralizing. While I understand as an artist you would be annoyed by this I feel that we need the art equivalent of a low-code/no-code tool and I'm glad we have one.
So? It's gonna be like when we built canals and disrupted the artisans and drove off the farmers from the countryside. It's simply the artists' turn now.
@@salazar1554 These aren't equivalent. That's a tool that helps. The equivalent here is "how to make a game without a developer."
Before even getting into the fact that game/web creation tools are products made by groups that are providing their knowledge of the fields knowingly/willingly.
Whereas AI to anything near its current state doesn't exist without massive art theft.
@@OatmealTheCrazy Okay that's fair actually.
Unless you have a genuinely creative mind then 'AI art' is like Photoshop on acid. It's borderline limitless already and it's only improving. But yeah, it's just a tool. It's only ever going to be as powerful as the craftsman. And that's what separates the wheat from the chaff :)
Who in their right mind would follow an AI "influencer" account?
Not much different from a real one given the depth of the content tbh
Dang, I'm getting too old for this 🤕😅
2:07 i need to know where i can find this image, pls
I like the memes that you integrate in video SOOOO much 😂
The way that you’ve pronounce FOoOoCus… oof 😂😂😂
Wow, dehumanization is closer than never before. People are amazed by artificiality and following fake people, whereas great artist like Martin Scorsese are booed for trying to keep alive some of the most difficult forms of art like film cinema
That's nothing new, people have been following fake celebrities and influencers since Instagram's creation. People who pretend to be rich by renting expensive cars for 1 hour to take 20 photos or going to fake private planes. Everyone is trying to blame AI for this, but humans were always like that. Past writers, such as Lovecraft, died in poverty and were only discovered after their death. Today, many talented musicians are unknown, but some wannabee gangster singing nonsense on autotune is filthy rich. Don't blame AI, blame humans.
I feel the best part of this is the accessibility to make unthinkable creations.
There's a channel called "There I Ruined It" that mashes up songs and vocals.
He made a mashup that had Dorothy from Wizard of Oz singing lyrics of Smash Mouth's hit "All Star" in the tune of "Somewhere Over The Rainbow".
Created on the news of the bands singer passing away, that is what this should be used for.
yeah but the way AI created content has been going, even memes wear out pretty quickly without genuinely funny human direction and creativity guiding it. Even all these anime girls who look "better" somehow look like nothing new and have the same soullessness. i don't know how else to explain it.
Even with these new tools, there will still be masterpieces that rise well above the rest.
@@rem7502 you know they're no real, that's all it takes, unless u live in the future like the weebs who think (replicant) slave owning will make a (public) comeback
but then again the dark possibilities can evaporate just as easily as every possibility. Weebs could evolve in a chamber of some sort, Cpt America style pod transformation
@@rem7502 The more art you see in a specific theme, the less special it gets. It’s normal, we all look in awe at things which is new to us, but we don’t care for the thing we experience day after day. Art will be (or more like it already is) mass produced without meaning behind it. There will be no message behind those pictures, they are superficial so you consume more and more without ever needing to use your brain. It mirrors our culture perfectly.
What an amazing video. Explained everything in such a friendly manner. Modt of the parts were funny bcz of the way you edited 😂
So, where can I download this all-in-one image generator program that has a user interface that basically works like Star Trek's Holodeck voice intup (but with text)?
I don't care about training the modules or whatever. There shas to be no more training to be done, then "I want to get this speciffic artstyle as a result.
But for everything else the program has to be in a usable stat already with release. Like I said, I dont care about individual modules or even training them to "my needs".
I'm not a Linux user, nor am I willing to become the equivalent of it just to be able to frankenstein my AI-tool together myself. Right after downloading and installing the program it has to "simply just work".
What you forgot to mention is that most of the SD1.5 anime models are actually basedon NovelAI's leak (animefinal-pruned) - and since the production of this video, NovelAI has released their v3 model which - as you might expect - is their anime SDXL model. It slaps, overall. They have some great training and technology that lets you prompt just about any character without a LoRA, its just in the base model.
Traditional and digital artists: We're going to be replaced with effortless AIGIs!
AIs: FREAKIN ROCKET SCIENCE
Hi :) Where can I find image from 2:09 ?
Where do you find all these memes! They are so funny lol.
Great video!
it's so not funny the more the image is "processed" by AI the more they look like the stolen copyrighted images the AI was "trained" with. a literal army of litigations against Open AI are in the works
What, they literally don't look like the "stolen" images the AI trained with. Idk where you anti-ai bros get all your goofy ass arguments from.
As a internet user for almost 20 years I can say that one thing is leading humanity to it's future: Porn
real
This is a simple thank you to bycloud's editor for explaining how to say Lycoris.
Great video! I am very late to the latest updates of AI gen
this AI thing's might be gettin traditional art to be favored over digital art again huh
?
AI still can't imitate the sense of purpose and my exact idea, so I decided to start learning drawing instead, and it's infinitely more satisfying than proompting.
That's because these apps are impressive but very narrow. AGI is a long way off. A book called The myth of artificial intelligence explains why. Good luck with your new art.
IMO using AI is an art form if you use it in all its complexity and strive to understand how it works, but it's a very different form of art compared to drawing or even editing images. It's like comparing digital music producers to people who want to learn physical instruments.
Just using prompts to generate images though? I'd say that's more of requesting art from the artist (which in this case is the people who have created and refined the software). Just like with actual drawing artists, you can't really get the same level of purpose and idea by commissioning someone to do it for you.
None of this to say I'd disagree with you on this in any way, I just wanted to share my thoughts on the subject. I don't really do image prompting because I'd like to learn it properly and I know I won't - which is the exact same reason I don't really do drawing either lol
based
@@Pawtacle wdym digital instruments are still instruments.
@@binyot5505 Sure, what I mean is that the tools you use are different even if the result is music - just like how AI tools are different from drawing tools but the result is still an image. It's a completely different skill set.
Are there a model specialized in space image ?
Ebank art looking real nice! Crazy how fast things are happening
The methods producing a high quality model are actually quite clear. The thing is that people usually looking for simple and quick methods, but quality requires a lot of prep work before the training even starts. And how much work that is for you depends on your skills and how you combine them to achieve the goal. And that is preparing the data set.
It is easy collecting huge amounts of pictures you can feed in your model, but that does not improve anything on its own. You need to carefully select pictures the "system" can understand and then also describe the pictures accordingly. And quality of the data set is another thing. Basically all the content you can find in the internet is fragmented jpg material, which needs to be cleaned before training and such things.
I have for my own work a data set with around 2200 pictures, and with that i mean exclusively the hand picked and tagged, partly extensively modified ones that are ready to be used in training. The actual amount of pictures i have, i don't even know really, it is more than a million, which i can go through and pick new ones to add to the data set. And over the last 6 months or so the quality of the whole thing has improved so much, that my older pics in the ready data set actually kind of meh by now. But i am at a point where my model is able to produce its own training material with the help of some tools and external input. The quality is so good.
But here is a thing i noticed about the super realistic models that gives a hint of the limitation of the authors training work. Especially those models that produce very good looking subjects very consistently are quite static, and to prompt them to do certain things with those subjects may require even external help like control net or other means. What i learn from that is that the input material has been kept in a certain likeness. For example people may have different styles of clothing and appearance but a basic pose. Something like that makes it much easier to process the input material before hand and you are able to have a large enough amount to make the "system" understand it well enough to output it consistently.
That is the whole trick really, nothing to be secretive about actually. I am actually working on braking the static and keep prompting simple yet versatile. And usually i don't use any negative prompts. Instead of using a Lora to improve hands i have around pictures i can train any model to improve hands. Its not perfect, but it is working and i am still learning to improve this process.
But i will say, i wont share my data set nor how i tag my stuff, because that is where the work is and where quality is decided. Why would anyone just share that? Except maybe this person is very wealthy or something? You can already download the model and even merge it with your own. But the foundation of that, the data behind is more valuable as you can use it any time again and improve it. The companies working with AI spend millions for that kind of work. The computation is kind of the easy part in terms of human involvement.
Not to be "that" guy, but do you have any right on the pictures you use ?
A lot of artists are strongly against their work being used by AI.
@@WallySketch Lets be frank about it, the essence of fair use is that using copy right material is legal as long as the result is transformative and not just a direct copy.
And I am a fan good privacy laws, because I think you should have a choice wether to be a public known figure or not. If someone puts their face in every camera I don't have any fucking remorse.
Maybe those artists can use their reach and help teaching young people to make a reasonable use with social media but I guess thats not the point why they care in the first place.
@@madrooky1398 There are others aspects that need to be regulated before it could become fair use.
Is it public ?
Is it commercial in nature and motivated by profit ?
Does it use the whole image, or only parts of it ?
Does it provide a significant market substitute for the protected aspects of the originals ?
As for arts on social media, it is often asked in my field (illustrator, comic book artist) to have an online gallery. We don't really have a choice, that's how we get jobs and pay the bills.
Nowadays everyone is looking for artists on the internet. If we're not showing our work, constantly searching for opportunities, and showing we are active, we don't get jobs.
So you are using an open source software for free using models that were shared for free with probably a shit load of extensions that were developed and shared for free. You probably read or watched a lot of tutorials shared by the community and used thousands of images you never bothered to pay. BUT you consider that your work is far too valuable to be shared. This is another level of being an asshole.