4:44 Midjourney does not use reactions to the images in production to train their model. It's a good example to explain it as a hypothetical, but it's untrue.
I still don't understand how Stable Diffusion works, but now I know more. Maybe you can help me understand what's happening when I try to create some art: First, I upload an image to Stable Diffusion in the img2img tab and then I select Interrogate CLIP or Interrogate DeepBooru, then I copy/paste the prompt into txt2img -- Why don't I get an image that better resembles what I started with? How can I get better semblance to my original image? You seem to understand this stuff better than me, so maybe you can explore this in a future video. Thanks!
such a good video, surpised to see so few likes. your explanation is great! since it works fine for a wider audience with minimal engineering or technical skills. please keep making the videos!
after watching the video of the video, I still have questions. It turns out that we make Gaussian noise from the picture, and then we make noise back from the noise. But won't we be able to face the fact that the noise can be the same?
Pretty unlikely if you use a random seed to generate the noise, and you train it 1000 times per image. The odds of getting the same noise that many times are mathematically improbable.
OK, but I have the same issue; it sounds like the system requires one to 1) upload a bunch of images, 2) reduce those images to "noise" and then 3) use prompts to bring the images back....makes no sense
Wow, such a great video man. Finally found the video that clearly explains how exactly images are made from text prompts. And the things you said in the end... yeah man... I agree with you. We should be careful on how to use these AI technologies.
@@allyourtechai let say the text prompt is “Rainnbow unicorn” . How does the process starts out ? Where does it get the noisy image of that in order to work back to the desired image?
You can have a completely offline install, where you download the checkpoint and other files, run the Stable Diffusion server on your own computer and control it from the browser on that same computer. No one ever looks at what you generate or charges you for anything. And you can train your own checkpoints or embeddings locally, but that is really slow (several hours for like 10-50 images and a RTX2060).
@@krzysztofczarnecki8238 i think thats what i got rn, its pretty cool running it locally. And yeah i pulled my internet plug and it was still able to draw somewhat accurate drawings of famous anime characters which is pretty awesome.
(9:08) Training the AI model with your custom data works better if you a) make sure each image is a square 512x512 pixels, and b) take the photos of your models specifically for this purpose in front of a solid color background. Also, I dare you to use "me from behind" in your prompts, as all of your photos appear to be selfies so it has no idea what the back of your head looks like.
You made a strong point on the confusion between really and AI generated really. Not to be pessimistic, but this is a huge risk for humanity. I believe right from the start we should have regulatory institutions to force AI companies to put a disclaimer on any art or content that’s produced. Tools should be developed and make It available to people maybe through their phones, laptops, tv as an extension so they can clearly differentiate between both. With the consumption of content being already high for most people, these technologies can easily turn into tools of mass control if strong measures are not taken right from the start.
People will just remove those things. I'm sure something will happen, the govt will probably step in, and do something stupid because they're all old, ill-informed, and don't understand how e-mail even works Whatever they do will either be over the top, or a waste of time. I think most people already know fake images, video and audio are already circulating. People are already questioning anything they see, so I'd say awareness is already out there. We just have to hope that "trusted" mediums don't mislead people with fake stuff, and actually do a little research. Fortunately (and unfortunately), I think most Americans already don't trust the media as it is right now. Especially with all the law suits these companies have had to pay out over the last few years
See. This is why my phone don't know what to do when we start stable diffusion Because the archives don't have this information It does,but it has to search in other places, and I don't know just how in the world I could do this on my phone??!! I don't think my " app" would have enough storage to do it. Or somethin somethin. Also,does my mother know her computer has a virus?? Just wondering if she knows that there has been a file downloaded to her computer and she might need to get it looked at by someone other than flight simulator or RAY saville. LMAO also,I guess American airlines don't allow this huh? Can we switch countries in tor and go to Singapore or something?? Lol wow. I'm not Anna btw. Y'all apparently have her in the pod...cast . This is how y'all make your dirty laundry money!! This right here!! Right?? Y'all do zoom calls during pandemics and get facial recognition and zoom call meetings and that's how you do your scams and make your dirty money. Wow Thanks for the info Did y'all do this to me????
You forget that this poses a huge problem for the legal system as well. Pictures or videos of you doing something are essentially worthless now given how easy it is to fake them.
jessus christ pliz mix yuour voice with some eq you have a terrible amount of sub bases (between 60 and 20 hz). Please ask your musician friend to show you how to do it bc it's unlistenable on many types of speakers
4:44 Midjourney does not use reactions to the images in production to train their model.
It's a good example to explain it as a hypothetical, but it's untrue.
It's criminal that this only has 13k. Keep it up!!
You should change the title, this is definitely not a "detailed explanation". It's more akin to a "summarized intuitive explanation".
Scott Morris works for the govt?? Lmao and Michael too?? And me?? Naw. I just call people like you out
I still don't understand how Stable Diffusion works, but now I know more. Maybe you can help me understand what's happening when I try to create some art: First, I upload an image to Stable Diffusion in the img2img tab and then I select Interrogate CLIP or Interrogate DeepBooru, then I copy/paste the prompt into txt2img -- Why don't I get an image that better resembles what I started with? How can I get better semblance to my original image? You seem to understand this stuff better than me, so maybe you can explore this in a future video. Thanks!
I will do a video on the subject. There are definitely some tricks to making it work and getting a decent result.
It just clicked at 4:38 why midjourney and others are free to start, they need people to teach the system
This was so informative! Thank you, love your videos!
Thank you so much, I really appreciate it!
Amazing explaination on such a short video !! Keep up the good work !!
Very well explained, thank you. And man, I love your studio! (D'oh - just noticed it is a fake background. Rather goes to your point).
Haha! You nailed it
such a good video, surpised to see so few likes. your explanation is great! since it works fine for a wider audience with minimal engineering or technical skills. please keep making the videos!
amazing. Thanks +wait for more ❤
after watching the video of the video, I still have questions. It turns out that we make Gaussian noise from the picture, and then we make noise back from the noise. But won't we be able to face the fact that the noise can be the same?
Pretty unlikely if you use a random seed to generate the noise, and you train it 1000 times per image. The odds of getting the same noise that many times are mathematically improbable.
But why is the "image" hidden in the "noise" to begin with? I kinda get what you're explaining, but I don't understand the starting point.
One basic question..What is the need to introduce noise in the first place?
The noise is the starting point when you reverse the diffusion process. It also provides randomness to the resulting image
OK, but I have the same issue; it sounds like the system requires one to 1) upload a bunch of images, 2) reduce those images to "noise" and then 3) use prompts to bring the images back....makes no sense
One of the best videos iv'e seen in a while. Thank you for taking the time and making such awesome content. Much appreciated.
Wow, thank you! So glad you enjoyed it
This was great. Thanks!
Wow, such a great video man. Finally found the video that clearly explains how exactly images are made from text prompts. And the things you said in the end... yeah man... I agree with you. We should be careful on how to use these AI technologies.
Very well done. Many thanks
Thank you!
Great video 👍 Subbed
Thanks for the sub!
So how does it know which pure noise image to use starting out with?
The software starts with a random number generator that is used as a seed to generate the noise.
@@allyourtechai let say the text prompt is “Rainnbow unicorn” . How does the process starts out ? Where does it get the noisy image of that in order to work back to the desired image?
Great explanation!!
Thank you!!
Awesome video !
Thank you!
Really awesome
Thank you!
does it pull stuff only from the checkpoints used or also online?
You can define the source. In my video about how to “ai yourself”, I provided my own photos to train the model.
You can have a completely offline install, where you download the checkpoint and other files, run the Stable Diffusion server on your own computer and control it from the browser on that same computer. No one ever looks at what you generate or charges you for anything. And you can train your own checkpoints or embeddings locally, but that is really slow (several hours for like 10-50 images and a RTX2060).
@@krzysztofczarnecki8238 i think thats what i got rn, its pretty cool running it locally. And yeah i pulled my internet plug and it was still able to draw somewhat accurate drawings of famous anime characters which is pretty awesome.
It is an Erlenmeyer flask, not a beaker. ;)
Thanks Walter White lol
(9:08) Training the AI model with your custom data works better if you a) make sure each image is a square 512x512 pixels, and b) take the photos of your models specifically for this purpose in front of a solid color background. Also, I dare you to use "me from behind" in your prompts, as all of your photos appear to be selfies so it has no idea what the back of your head looks like.
Breeze?? That black guy at Walmart all crazy looking at me cause I didn't have on a mask - you know what's up don't ya?
You made a strong point on the confusion between really and AI generated really. Not to be pessimistic, but this is a huge risk for humanity. I believe right from the start we should have regulatory institutions to force AI companies to put a disclaimer on any art or content that’s produced. Tools should be developed and make It available to people maybe through their phones, laptops, tv as an extension so they can clearly differentiate between both.
With the consumption of content being already high for most people, these technologies can easily turn into tools of mass control if strong measures are not taken right from the start.
It’s something we need to all pay close attention to for sure.
People will just remove those things.
I'm sure something will happen, the govt will probably step in, and do something stupid because they're all old, ill-informed, and don't understand how e-mail even works
Whatever they do will either be over the top, or a waste of time.
I think most people already know fake images, video and audio are already circulating. People are already questioning anything they see, so I'd say awareness is already out there.
We just have to hope that "trusted" mediums don't mislead people with fake stuff, and actually do a little research. Fortunately (and unfortunately), I think most Americans already don't trust the media as it is right now. Especially with all the law suits these companies have had to pay out over the last few years
I like my noise Gaazian
who doesn't really :)
See. This is why my phone don't know what to do when we start stable diffusion
Because the archives don't have this information
It does,but it has to search in other places, and I don't know just how in the world I could do this on my phone??!! I don't think my " app" would have enough storage to do it. Or somethin somethin. Also,does my mother know her computer has a virus?? Just wondering if she knows that there has been a file downloaded to her computer and she might need to get it looked at by someone other than flight simulator or RAY saville. LMAO also,I guess American airlines don't allow this huh? Can we switch countries in tor and go to Singapore or something?? Lol wow. I'm not Anna btw. Y'all apparently have her in the pod...cast . This is how y'all make your dirty laundry money!! This right here!! Right?? Y'all do zoom calls during pandemics and get facial recognition and zoom call meetings and that's how you do your scams and make your dirty money. Wow
Thanks for the info
Did y'all do this to me????
"my hope is that it brings us all closer together..." yyyeeaaaa....that's a no from me dawg
Oh shit it's you brina 😂
Haha! How have you been?
You forget that this poses a huge problem for the legal system as well. Pictures or videos of you doing something are essentially worthless now given how easy it is to fake them.
I have no idea who Drake is
Well now you do hopefully!
jessus christ pliz mix yuour voice with some eq you have a terrible amount of sub bases (between 60 and 20 hz). Please ask your musician friend to show you how to do it bc it's unlistenable on many types of speakers
Youre great