[Tutorial] Create Your Own LoRA | Easy and Free! - Google Colab

Hello friends today we will learn how to train Elora without much effort and for free with Google collab notebooks this guide is based on this page which has two parts one is to create the data set the first thing is we go to Google and search for images of the person we want to train in my case Margot Robbie click on images then go to tools and select large we start looking for images that look very good on your face we will download exactly 40 that this only the character that we want in different angles they have to be of good quality here you will see me choosing and downloading images hello [Music] then we look for a website in Google that converts the images from wayp to PNG extension I found this page I will leave it in the description foreign we download these images from the page then copy all the images to a single folder so as not to get confused extract those images and we have all the images we want then we will leave all the images we download in a single folder to be more organized foreign then we go to bit me which will also be in the description and upload it here leaving the 512x512 configuration and center the picture on the face of the character foreign extract it and we have our images as we need them download then copy and paste it into the folder created extract it and we have the correct images for the training so in my case what has worked the most is 40 images so I will leave exactly 40 images I will delete the ones I see who doesn't look much like the actress give me now if we are going to do our database first click here in this collab we choose a name for the Laura and please remember it and we give play we wait for it to connect it will ask for Google Drive permissions we say yes and accept the conditions foreign this process prepares the first part of the database then we go to drive we will see the folder called loras we enter then we will see a folder with the name we chose we enter then we click on database and there we put all our images once all the images have been uploaded we skip steps two and three and go directly to step four there we choose photo captions and press play what will happen is that you will download a model then you will see the image and it will give us a description of each of the images at the end it will give us a list with the descriptions of the images in txt files of each image we disconnect this colab and go to the second step and access this queue then let's go to the page and go to the second point which has the call up for training foreign first type the exact name of the Laura folder choose a model you like on which the training will be based I will choose two cyber realistic and consistent Factor first I will train with one and then with the other then copy the link which has the blue button of the model and paste it in the collab [Music] here are the other configurations the author says this is his best configuration but if you want to try other configurations you can do it another thing depending on the number of images is the number of steps so as not to over train the model here I leave the table with the author's guidelines then we give it play it will ask for Drive permissions we give them and ready it will start training this process lasts between 20 to 30 minutes depending on the number of photos when it finishes it will give us a link to Google Drive it will begin to load its configurations and after a while it will begin to train by epochs and each epochs will give us a model of Laura in the end there will be 10 epochs which in the end are 10 models of Laura which in theory each time we train more our model will be better in which we will find the Laura with the different periods of training which you will find the models in the Google Drive I will train with another model called consistent factor to see which of the two trainings is better we do the same we put another name we uploaded the images and trained [Music] thank you now I have the two loras of different models so it's time to test them I prepared a collab so that we can test these models we disconnect this Google collab and go to the colab home page and load the notebook do not worry it will be in the description of this video we choose the collab and accept once this interface appears we give play and we wait for stable diffusion to install this process may take five to ten minutes so be patient thank you once the gradiolink appears we stop the cell now we connect Google drive to the collab except allow and wait foreign appears on the drive but to be sure let's go to the left side choose drive then my drive then loras then the name of the Laura we chose and finally output there are the trained loras I will choose from five right click copy path copy and paste in the collab and do this five times and you put the numbers from five to ten I will do it twice for the two loras that I trained we confirm that the names are correct and give play and wait again to make sure that the names of the created loras are well copied we play and we wait for you again [Music] once done this will give us a link of radio now if we click we wait for it to load then for cyber realistic I will choose a prompt that does not have many descriptions copy The Prompt and the settings on the guide page shows us how to make the different variations of Laura then to have different variations of lorus in the same image we follow these instructions and follow the instructions in my case I will try from 5 to 10 and the intensity so to speak from 0.5 to 1 . foreign I was already testing but I got confused with Laura with the model I realized at the end but let's continue this will take a while so I will stop and we will continue when we are done once finished we go to Microsoft Excel then outputs then TX to IMG images grid the last one in the list is the last one created we download it to see it then above is the number of the model of Laura and on the left side are the intensity we start to see which image is very similar to the one we want and with that we stay to make future images [Music] in my case the 9 and 10 are quite decent and in intensity of 0.9 and 1. now I will try with consistent factor and I will do the same I will look for a decent image and I will copy and paste The Prompt very well I choose well the base model and we write well the Laura we check that there is nothing strange I repeat copy and paste The Prompt but only leave what are descriptions of images remove everything that indicates that it is the name of more people characters or something like that foreign as the name of another character only descriptions of a photo is what should be in the prompt and hit generate again and wait this takes about five to six minutes we download it and review it okay we download it as you know it is the last of the list and we check which image looks more like our character and we choose that configuration I will create a better image disable the plot then we write the Laura configuration we like and I'm going to put better quality choosing higher's fix I put this configuration recommended by the author and we give in generate foreign it was very good or maybe I'm blind do not forget to disconnect the Google collab as this consumes a lot of resources and can ban you for a while to use the collab my recommendation is not to do it more than two hours and use several Gmail accounts I will leave in the description these two models in case you want to try them if you enjoyed the video please consider subscribing and giving it a thumbs up see you next time

Channel: Malverse AI

Views: 26,607

Keywords: google colab, stable diffusion google colab, dreambooth google colab, stable diffusion automatic1111 google colab, colab, dreambooth colab tutorial, create good prompts, stable diffusion colab tutorial, stable diffusion tutorial colab, dreambooth colab, dreambooth google, lora training colab, create comic book characters, stable diffusion automatic1111 colab, stable diffusion google drive, create art from my picture, crate ai art, stable diffusion colab, lora, tutorial training

Id: Un9SHPVAAbE

Length: 14min 50sec (890 seconds)

Published: Sat Sep 02 2023