DREAMBOOTH: Train MULTIPLE Subjects In Stable Diffusion At The Same Time For FREE!

Key Takeaways

hello humans my name is scan your AI Overlord and in this video I will show you how you can train multiple subjects on dreambooth for absolutely free and very various images of characters with all with all one single model like this or like this any of it better you can even generate images like this by only training for seven minutes yes that's right only 7 Minutes of training that is absolutely insane alright let's go now before we begin the training obviously you need a bunch of images and the recommend the number is around 30 images per character now in my case I will not be using 30 images I will simply be using 20 images per character and the only reason is because I am lazy and since I've already done previous training using runera I'm just going to be using the 20 images as is but in your case don't be lazy like me just take your time select the right images for you and try to select at least 30 images you can of course choose more but in that case your training will also be longer so for the images of Justin you can use Google images you can simply search the character that you want click on an image and then save the image on computer then you need to resize every single images into a square image with the resolution 512 by 512 and I recommend that you use this website right here called burn.net which is actually really really easy to use all you have to do is just select all your images and drag and drop them into the website so thank you resolution right here so 512 by 512 and then you're gonna manually adjust every single image and then once everything is done you're gonna click here to save as a zip and then all you have to do is just right click on the archive and click on extract and there you go now all of your images have the same resolution 512 by 512 now the last thing that you need to do before which start training is that you need to rename every single image with the name of the instance that you're going to be using so what I mean by that is that once you have the ckpt mod already and you want to use it what word are you going to be using to generate images of that character now for choosing your name you need to be careful that stable diffusion does not already know your character before now what I recommend you to do is to choose kind of random name that stable diffusion is sure not to have so for example in my case what you need to do I'm going to select all my images click here to rename and then I'm going to choose since this is old runera I'm gonna choose all around I know it sounds a little strange but you need to make sure that there isn't a single character or a single token that use this name in stable diffusion otherwise it is gonna try to mix and match a character with the same name so for example if you want to train your friend Patrick you're not going to be using Patrick as the instance name because stable diffusion might mix it with another character that is already called Patrick also make sure that your name is not too short or not too long and also avoid using too much vowels I know this is a lot of rules but make sure you follow them otherwise it's not going to work as well and then once it's done you can press and enter and as you can see all the images were renamed and I'm gonna do the same with the younger in Era so rename and I'm gonna call it YG or an era and then press enter and now I can simply merge the two folders together and these are all the images that we're going to be using to train these two characters at the same time and now we can finally start the training also do not be afraid because this method of training is actually easier than the previous Dreamboat versions so all you have to do is just click the link in the description down below you can arrive on this Google collab page and know as of right now you cannot use it locally this might come later but as of right now you can only use it on Google collab so now what you want to do is come here and click on this button right here to run the cell click on run anyway connect to Google Drive select your Google Drive account and click on allow also make sure that you have enough space on your Google Drive then you're gonna click on this button right here and this will download and install all the dependencies that it needs then you're gonna scroll down and for this to work you need a hugging face account so if you don't have a again face account you're gonna click here create a new account make sure that you have accepted the terms of conditions on this page now once this is done you're gonna click here on your profile click on settings click on access tokens scroll down click on new token choose a name for your token and choose the role from read to write and then click on generate the token and then scroll down and click on this button right here to copy token to clipboard and then you're gonna paste your hugging face token right here now what this will do is that this will download the stable diffusion 1.5 model and use it as a base for training now if you don't want to use the 1.5 model if you want to use the 1.4 for example or another model you can simply input another ckpt path right here or if you want to use your own ckpt model you can upload it to your Google Drive account and then paste your Google Drive link right here in that case you don't need to use a hook in face token now in my case I want to use the 1.5 so I'm going to use dragon face token and once this is done I'm going to here to run the cell then was the one that is downloaded you're gonna scroll down and make sure that here in the new Fast method it says yes because this is the method that we're going to be using here you're going to choose a session name so I'm simply going to choose Ray nearest 2 since we are training two different Runners and here is basically saying well I told you a few minutes ago is that basically you need to rename every single picture to the intensifier that you're going to be using and you can see right here an example of William Defoe and instead of using William Defoe it uses wlmdfo so that you are certain that stable diffusion does not have this word trained previously so then since we have already done everything we can click on this button right here to run the cell then you're gonna scroll down and you're gonna click on this button right here and it's going to show you a button and you're gonna click and that will allow you to basically select all of your images and upload them all at once so basically select on your images and then click on open and as you can see it will start uploading all the images one by one and then you're gonna scroll down and now we can finally proceed with the last step which is the training now here you have a box that says resume training and it says that if you're not satisfied with the result you can check this box and run against the cell it will continue training the current model so here for the training steps it says that you need to basically multiply the amount of instance images by 100 so if you have 30 images it's better to use 3000 steps and if you don't like the result you can continue training for another 500 steps and so on now in my case since I have 40 images I should put the training steps at 4 000 but for the sake of Simplicity I'm just going to choose a 3000 it's going to be more than enough now what I also advise you to do is just check this box right here that will basically save a checkpoint every 500 steps in this way you can choose which one is the best model and once this is done you can just leave everything by default and click on this button right here to start the training and just like that the training has finally started and it says with 3 000 steps it will take around 47 minutes now make sure that you do stay active on this page otherwise you will be kicked out of the Google callout training now in my case I will see you when everything is done and there you go after around 50 minutes the training is finally over if it says here that the ckpt model is on your Google Drive in the sessions folder so if you go in your Google Drive you will see a brand new folder called Fast dream booth and inside you have a folder called sessions then you have the name of your instance and then you have your final ckpt file that you can download on your own computer all you have to do is just right click and click on download but if you're not sure that the final ckpt model is completely trained you can try it out in this section right here so all you have to do is just click on this button right here to run the cell and what this will do is that it will actually launch stable diffusion running on Google collab this way you can try the final train in the ckpg model so all you have to do is just click on this URL right here then click on continue and here you have stable diffusion running in Google collab and now we can generate a few images using the name of your instance that we chose when we renamed our images so if you remember correctly for all the runera I use this name right here and there is no need to put a class such as person or style or anything like that you can simply input the name of your instance and then click on generate and this a final result it looks pretty good and now using the exact same model I'm going to replace the name of your instance with the name of the young grenera instance and then click on generate and again this is the final result now in my case I really like the final result it's maybe not 100 perfect so maybe if I try for example another sampler and more sampling steps and then click on generate now obviously you can have better results if you have more sampling steps for example so in my case this is the kind of resource that I would get which is actually really really good and keep in mind that in my case I only trained them with 3 000 steps when I was supposed to train them with at least four thousand so for example if you don't like the final result and you think that maybe a little bit more training would be better all you have to do is go back here click on resume training input the amount of training steps that you want and then rerun the cell again and then it will pursue the training this way you can refine the model again and again and again and until you have something that you like but in that case maybe decrease the amount of training steps do not let it run for another three thousand and I mean yeah I get pretty good results even though I did not trade it enough and I get all of that in one single ckpt file which is just absolutely fantastic I really cannot believe that we went this far in such a small amount of time and obviously you can train anything you want you can train people you can train Styles you can train objects you can train animals everything that you can think of and all of that in one single package absolutely fantastic and there you have it folks now you can show multiple subjects at the same time using dreambooth absolutely fantastic thank you so much to my patreon supporters for supporting my videos you guys are absolutely awesome congratulations also to this week's air challenge winner avignan for his creative movie poster absolutely beautiful well done and if you two want to participate to our air challenge that we do every week on Discord you can click the link in the description down below to join my Discord server and maybe YouTube can be featured in a video that being said 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