00:00Everyone, welcome to another AI Fuzz video
00:02My name is Marcus, and we are here to cover a huge, huge node for ComfyUI users
00:11Yes, you can now train fully trained Loras exclusively in ComfyUI
00:20You heard that right—no more Civitai, no more Kaggle, no more Google Colab
00:25You do all the training in ComfyUI with just one single node
00:37Yo, let's do this! Alright, so first, we have—I'm going to put the link to the GitHub for this node
00:43It is Lora Training by LarryJane491
00:49I'll link it in the description
00:55He also has a group of nodes and can create caption text files for each of the images
01:02 So, this is how it starts
01:09You're going to start with creating a dataset—right? You need a dataset, a dataset of images because
01:16That's all it is
01:17The dataset is just a collection of images that you will train AI on
01:23So, here I have a group of sketches that we are going to make a sketch style from, and this is all just a group of sketches
01:34I usually get maybe 25 to 50 images for this demonstration; we're doing 24
01:41But these are all different types of sketches, each in their own PNG format
01:47They don't have to be the same size, and you put them all into a folder designated as a database folder
01:53It's very important to note that you have to name this folder, for example, "folder five" or more, and the name of your LoRA, whatever you're doing
02:05That folder needs to be inside a database folder because that's the structure this node works off of
02:13Once you have your dataset in there, first we're going to make the text
02:19Captions for that
02:25So, what I did was I have a brand new, fresh install of ComfyUI because you need to have a certain version of Scorch CU 1.21 to work
02:31If you try not to, it's going to break
02:36So, I have a separate ComfyUI installed that strictly has that version on it, and it's strictly for training LoRAs for me
02:46All right, so once you download the nodes from GitHub, it's going to be right here, Lora
02:54And you're going to have big nodes here
03:00So, to do the LoRA caption save, we're going to do the LoRA caption load
03:06Okay, and in the middle, we're going to do the WB 1.4 tagger, which you can download from the manager as a plugin
03:19Just load it and type in "WG." It's right here, so just download that
03:27We start ComfyUI, and you'll have it
03:30Here is a simple process
03:33All you have to do is plug the image list into the image input here and connect the string through the text prompt
03:43Then, plug the path into the path input
03:46Hoops to the path, leave the main list open, and then we need to do the pick it in here
03:52Hit OK
03:53Default settings are fine for the width and height, and all you do is prompt
03:58Oh, I forgot to view the name list, and then you prompt
04:02I guess you do have to include the name list, sorry
04:06So, it's GNA go in
04:09Basically, now it will analyze each image in your dataset
04:12It's going to create a text file that describes what's in each of the pictures, so when it trains, it kind of knows what’s in each image, giving it some idea of what it's looking at
04:30Okay, so it's all done
04:32Training, and if you look here, it shows some of the text, the white IDs, and some of the photos you can use
04:44You can really use any of these as trigger words for your LoRA
04:48I usually just use the name of the LoRA as a trigger, and that always works
04:53So, if we go into that file, the database file, we will see now there are text files for each of these pictures
05:07Do that, all right? So, the text files are there along with the corresponding images
05:08Now, we have to bring up the Magic Node
05:13So, let's clear this, put it in, go to node LJR, then back in
05:19And this is the Magic Node
05:24It's for training in ComfyUI
05:26Huge! I remember Epic was telling me this is the biggest thing she's ever seen
05:33You love VFI and he uses it all the time
05:36So, ComfyUI, these are the..
05:40Options here: you have your checkpoint name, which is the base checkpoint you want to train on
05:46You have the path to your images, your dataset, batch size, MAX epochs, save every EPOCH
05:56The output name is what you want to save the output as, like "Loras" with the name you choose, and the output directory
06:06So, we are going to do this as we did before; let's specify the path to the folder
06:14Now, this is very important: for this, you don't put the full folder path, just the name of the folder it is in
06:29Just "G", the database, and it will go into the database folder and look at that folder to automatically know where the dataset is
06:39Keep the batch size at 1
06:40You can do anything between 40 and..
06:44Above is fine
06:48I do 50 and save every an epoch, essentially
06:52What it's going to do is go through 10 images, then save an epoch, save a Lora from those images, then it will go to another 10, so they have 20 images to train on, and it'll save a Lora at that step and keep going—10, 10, 10—as to get to 50
07:10You'll have a full Lora with no, you'll see it with no numbers in the name
07:15It’s just going to say whatever the name of the Lora is, say Teniser
07:18Okay, the desired name
07:22I'm just going to name it SK because that's what the folder is named after
07:27You can name that whatever you want
07:29Clip skip will keep it at 2, and the output directory will be your Models Lora
07:34So, it's actually going to put it into your Models Lora folder for you
07:38This is..
07:41This is ridiculous
07:45All right, so while that's done, all you do is input the prompt
07:53What prompt? That's all you do
08:00So if we go into here and look at what's going on, it's going to go through, find the values, and it's going to find the buckets with the images
08:08Then it's going to make the buckets and list the steps it takes to make this Lora
08:12Now, I usually keep this up while it does it, so I will keep it up for you
08:19But that's all — training took, whoops, 23 minutes and 20 seconds
08:29A trained model
08:39So then, a little basic workflow here: upload the Lora and set of these clips, and there it is
08:42You will do..
08:46uh, the prompt
08:49The prompt was, um..
08:52what do..
08:56uh, s, ah, all right, well, C that prompt
08:59So this is the prompt with the brand new SketchMo that we just trained
09:04Made trying to directly input into ComfyUI, trying to find a direct way of doing it
09:11Well, uh, let's do another one
09:16Let's do four this time
09:22All right, there goes the Treeline Street sketch
09:27I'm using RF Brandy sketch, Laura, and again, you can really use whatever you want, as long as you have the right images in her dataset and the training goes well
09:41And again, I did not use any Kya, no, no Kaggle, no Google Colab, nothing outside of ComfyUI
09:48It was all done within ComfyUI, and there is the output right there
09:53So, again, like I said, I will put the link in the description, and I hope you guys like videos like this
10:02I will show you a few images at the end of some trainers that Laura did prior to him leaving
10:08So, we'll see you in a few days
10:11Another AI
10:13The video features Markus out, and the crew will be back
10:17[Music]