DREAMBOOTH LOCAL Training Inside Stable Diffusion! CPU OPTION For FREE!

Key Takeaways

hello humans my name is overload and finally my brothers Rejoice because finally this is the day we finally have a dream Booth training in stable diffusion no need for Google call up docs or GPU ranking you can train your own images locally without leaving stable diffusion and that is absolutely fantastic are you ready let's go alright so before we begin we need to make sure that you are using the latest version of stable diffusion so for this you have two solution either you come here and you click on your photo URL and you type CMD and you press enter here in the command prompt window you're gonna type get pull and press enter and we saw update your stable diffusion folder the second option is to right click on the web UI user.bot file click on edit with notepad and here right above called Web ui.bat you're gonna type git pull and then You're Gonna Save the file this way each time you launch table diffusion it will automatically update the folder and then you can launch stable diffusion now before we begin I think it's important for me to tell you that if you do have a very a weak GPU this is probably not gonna work for you obviously and to be able to run this correctly you need at least 12 gigabytes of vram and even with 12 gigabytes is probably not going to be enough if you want to have all the bells and whistles so unless you have a very very powerful GPU this is probably not for you now as some of you know I also don't have a very powerful GPU since I have a 1080 with only 10 gigabytes of vram so to show me how this entire thing work I will basically be simulating a local environment using run pod so the entire training will actually be done by a very powerful GPU but don't worry the same exact steps will still apply to you if you do have a GPU that can handle this kind of operation okay so now to active in Dreamboat you're gonna click here on the extensions tab you're gonna click here on available click on this button right here load from and you're gonna see a bunch of options right here also if you're interested in a separate video where I show you an explain what are the best extensions available for a stable diffusion you can let me know in the comments down below but today the extensions that we want is right here the dreambooth one also if you click here you can arrive on the GitHub repository also if you have any issue with the dreamboof extension you can come here and search if other people didn't have the same problem as you which is probably the best way to solve your problem as fast as possible so to install this extension you're gonna click on this button right here and click on install and then click on this button right here that says apply and restart UI and as you can see we have now a brand new dreambooth tab also if this doesn't work for you I highly suggest that you actually restart completely stable diffusion because sometimes not all the requirements are installed correctly also if you relaunch stable diffusion and you see the same right here this basically means that you need to install stable diffusion in a folder that doesn't have any spaces as you can see right here I have a folder that's called super stable diffusion 2.0 that has basically two spaces right here now in my case I don't really care since I will not be using my PC to do the training but if you do plan on using this extension this is something that you need to keep in mind so then once you are inside the dreambooth tab we're gonna start by creating an empty model now we need to create an empty model before we actually start the training now I know this might be a little confusing which is why I named this extension super stable confusion because yesterday I actually spent an entire day trying to make this work until finally succeeding so this really took me a lot of time but I did it so you don't have to now the training that I will be doing today is again with the renera actress and the reason why I'm doing this is because since I've done this training multiple times with different dream Booth options I know what kind of results I can actually expect from this so here you're gonna start by creating a name for your model so in my case I decided to call it renira ddim SSC DCIM for the scheduler that we're going to be using an ssce for super stable confusion here you're going to select the source to your checkpoint you can either choose the 1.4 version or the 1.5 so for this I will be using the simple 1.5 version and here for the scheduler I highly suggest that you choose ddim this is the one that creates the best results and then you can click on the button create and there you go after around one or two minutes your empty model is finally created and now you can click on the train model Tab and here you're gonna see a bunch of options now I know this might sound a little complicated but this is actually not that difficult because there is only a few options that you really need to change now here the concept list this is something that you don't want to choose because for this you need to create a separate Json file but we don't need this because we're gonna start completing the rest of the options so here the instance prompt this is basically a short description of your subject using a unique keyword and and also the name of a class so if you've done previous Dreamboat training before you know that we need to use a certain keyword for the model to activate and you will usually use the name of your character for this so in my case it will simply be using the name of my model so renera ddim SSC and also the name of my class and if you remember my previous dreamboof video we know that we usually use the class person also if we take a look at our sample images we see that this is basically a photo of renera this is a photo of a character this is a photo of a person which is why here for the instance prompt we're simply going to describe what our sample images look like so in my case my images are simply a photo of character name person so if for example you want to train it on your dog for example for the instance prop you will probably use photo of name of your dog dog instead of person and of course you can use this for objects in the style too so instead of using person you can use objects you can use style and you can use anything that you want as long as this represents the class of your subject so the class prompt is very similar to the instance prompt but here we just don't use the name of the character we simply input what our subject actually is so in my case this is simply photo of a person simple as that so again if you want to use a dog for example if you want to train a dog you would simply use here photo of a dog so just replace person by the class of your subject here the data set directory is the folder of where your simple images are located so for me my images are right here so I'm just gonna Ctrl C Ctrl V so the classification data set directory this is where you would put the regularization images now I know this is optional but I highly suggest that you actually use it because if you do use these regularization images that means the that you do activate the prior preservation option which is essential if you want to have the best results so for me my regularization images are right here and will also leave a link in the description down below where you can download this entire folder these are again the same 1500 images from the Joe Pena repository that we use the in the first dream book video so I'm just gonna come here copy the folder URL and paste it right here total number of classification images to use so as I said we have here 1500 images now if you don't want to download these images you can actually input an empty folder and here you can input the number of images that you want stable the fusion to create on the spot before it starts the training and usually the best number that you want is at least 10 times the amount of sample images that you have so let's say you have 20 sample images the total number of classification images that you want to create should be at least 200. now I would personally suggest just download these images this way you don't have to waste time trying to create hundreds of images from scratch I mean why waste time when everything is already done for you now here we have finally the training steps and the learning rate these are one of the most important options right here and you should definitely change this because the default ones are absolutely terrible they do not work well with each other so for this you have basically two options either you keep a Higher Learning rate in a low training steps or you keep a high training steps and a low learning rate otherwise you're gonna end up over training your model and you're not gonna have the results that you want now what I personally recommend you to do which is basically how I got the best results myself is basically a low learning rate of one exponent -6 which basically means five zeros right here and a training step between 1100 and 1200 these are the options that I recommend you to keep deep to really get the best results out of your training now in my case just for fun I'm just gonna put 11 11. it should be more than enough then the batch size in class batch size you don't want to touch that you leave this at default same resolution now here for the save a checkpoint every end steps and generate a preview image every end steps you absolutely need to change these two values because as of right now this is completely broken if you leave it like that you will actually never finish and you will always get an error during training so what we recommend you to do for the save a checkpoint every end steps you're gonna input a value that is higher than your training steps so in my case since my training steps is 11 11 you can simply put 5000 here and 5000 here this way this option will never activate so then you can scroll down and leave everything by default you don't need to touch anything then you're gonna click on this Arrow right here that says Advanced and here you're also gonna have a bunch of options now here surprising enough you actually have an option to train using your CPU but I'm gonna tell you right now never use this option this is completely useless this is only for fun and Giggles because you don't want to spend 20 30 hours waiting for your training to be over this is absolutely not worth it so basically here the only options that you really need to change is right here use 8bit to decrease the amount of vram that it will use at the cost of speed and here for mixed Precision you need to change it to fe16 this is the only option that you need to change everything else you need to leave by default because all these options right here are checked so that you actually use less vram at the cost of speed which is obviously what you want and if you actually don't leave this checked even a 24 gigabyte SV card will not be enough and again all the oceans that we see right here you need to leave them by default and once you are done once you have completed everything do not forget to come right here and select your model so for me it's runera ddim SSC and then once you're ready you can click on this button right here to start the training and there you go for me the training has finally finished as I said I'm currently running this on run pod and with a 3090 11 training steps actually took me less than 15 minutes which is actually super super fast and then once the training system complete the finished model should actually be in the folders models stable diffusion and your finished trained model should be right here and for the model to appear correctly in stable diffusion you can click on this button right here to refresh the list and then you can select your finished model that is ready to be used and as an example to prove that it worked I'm just gonna put a prompt right here so close the portrait of Arena ddim SSC person as you remember these are the two words that you need to use in order for the model to work which is the name of your character plus the name of your class and then I'm gonna click on generate and this is the final result looking really really good I gotta say this is actually the best model that I did using the young greenera images which is also coincidentally one of the fastest dream Booth training I did to this day because this only took less than 15 minutes so if you do have a powerful GPU like a 3090 for example you should definitely try to use extension because this is a game changer and there you have it folks now you can train your own images using dreambooth without leaving your stable diffusion UI what a time to be alive and there you go thank you so much to my patreon supporters for supporting my videos you guys are absolutely awesome thank you so much for watching don't forget to subscribe and smash the like button for the YouTube algorithm and I'll see you guys next time bye bye