Fast and EASY Flux LORA Training with AI-Toolkit and MimicPC

if you're an AI art generation geek like myself then you've been watching the flux model I'm sure with great appreciation as we all have even despite some of its little shortcomings here and there and you know what's really hot right now is getting your own face into these things via lauras stands for low rank adaptation and it's just a way for you to tweak an existing model to include your information in there whether it's your face or whether it's your cat or your friend's face or an object or a style it doesn't matter you can use luras to customize the flux models if you have the tools tools and know how now if you're like me you've looked at a bunch of videos about how to do that and you've seen all sorts of solutions some run locally and those videos will take you about an hour to get through when they talk about how to install all the software and oh make sure you have the right GPU and all that other stuff and even the ones that talk about how to set up various softwares on remote servers it's still a process to go through I'd like to share way with you that you can create a custom Laura in about an hour for about $119 cents over at mimic PC you don't have to run your own GPU and in fact you can train multiple lauras at once hey it's not your computer it's not your system it's not your resources being eaten up I've tried several of these Solutions myself including ones that I've installed locally cuz I'm a geek that way but man oh man it just took up a ton of resources I couldn't do anything else on my computer and this way I'm going to show you it really truly couldn't be simpler you don't have to mess with settings or anything the default settings just as they are are going to give you great results why don't I stop talking about it and just show you this will of course require that you have an account over at mimic PC but having a mimic PC account allows you to run the latest creative AI software on another machine without having to have any experience on setting things up I've shared a couple of videos about mimic PC in the past and I'm thrilled to share that they're going to be a regular sponsor over the next few weeks so I'm going to be featuring a lot of the tools that they have here and show you how to use them especially for those of you who do not have powerful computer systems of your own but would like to get started in AI so we're going to start with the fun stuff creating custom luras of your face or a pet and showing you how to use those lauras either on mimic PC or on your own machine the most important element in any kind of training like this is the Source images they've got to be of high quality and they've got to have the right variety for the model to be able to be flexible enough to give you creative options when you go to use these models I ran several tests in preparation for this video and the first test I ran was on my face using 30 images I've used these very images like a year ago to create stable diffusion 1. 5 models and I already know what's wrong with these images but I used them anyway for this test to see what would happen but let me give you a heads up first of all you don't need 30 go with 20 make sure they are various angles various lighting various clothing various scenarios otherwise your model will be too locked in and you won't have a lot of flexibility you also want to avoid shots like this where my head is way up close to the lens and it distorts it this is another example I'm way too close this is another example I'm way too close the lens Distortion is just weird and this too too close it looks distorted what's going to happen is that the model learns that this is what my face looks like distorted so I'm going to get varied results but don't worry I learn my lesson as we go along and by the end of this video you're going to know exactly what to do but I'm telling you right now 20 images is good because I did so many models with the 1. 5 stable diffusion model are at 512 x 512 resolution and I've tested these and they work great but you can also use 1024x 1024 and of course they'll be that much better but I got to say I'm not at all disappointed with the results I'm getting with my existing 512x 512 images once you've got all your images you're ready to go to mimic PC and launch AI toolkit once you've set up an account over at mimic PC you're just going to log in and you're going to be taken to the screen where you have all of your apps now just to show you these are some of the apps that you can run here automatic 1111 is a very common stable diffusion interface for those who are interested in AI art and animation RVC is for voice cloning AMA allows you to interact with large language models it just goes on and on face Fusion for face swapping Coya SS is another model training tool fukus comfy UI chats all sorts of tools in here that you can play with without having to learn anything about configuring them today we're going to use AI toolkit to train these luras and when you see how simple it is you're going to be like I can't believe I wasted all this time trying to figure out other alternatives to get started you're just going to click on quick start and I'm going to recommend that you choose the large Pro computer you're going to have 24 GB of RAM I've tested this it is twice as fast as going with the large model for a dollar so you're for 19 cents more per hour you're getting double the speed those were my results on more than one occasion now because this is a model that's trained and it takes about an hour you're going to want to make sure that you don't run out of time or you set it up to automatically extend your time as you need to so that's what I like to do I just do an automatic extension and it says the mission time will automatically be extended by 30 minutes before expir expiration automatically but I'm going to recommend that you just set a timer and you know in about an hour almost exactly an hour it's going to be ready for you with you if you use this large Pro machine so once you've chosen that you just click on Create and start and in a couple of minutes the machine will be set up you don't get charged for any of this setup time okay here's the interface for AI toolkit and it really is very simple to get started with your lore you're going to name your lore you want it to be a unique name so in my case I'm just going to call it Bob Doyle and then the trigger word will also be Bob Doyle and then for upload your images I can just click and drag all these images in here let them upload and now we have the option to caption all of these images now I had heard that with people it's not as important to do the captioning for flux lores for whatever reason so I ran this first test without doing it and it worked fine as you'll see so for now we'll skip this part but in a moment I'm going to give you an example where I do use the captioning it gives you an example test prompt a person in a bustling Cafe and then Bob Doyle that's the trigger word I gave it up here so whenever it sees Bob Doyle the model is going to produce some sort of version of my face and that's it you just click start training it's going to tell you that the training has started locally and your Laura will only be available locally because you didn't log in with a right token to hugging face we want things locally because we're going to download them and use them on our own machine if we do that but we're also going to want to have them downloaded so we can upload them to another instance of a machine here on mimic PC we can check the progress by clicking logs Tab and it will show you what's going on here this quantizing Transformer if you're using the large Pro this step right here is going to take you about 3 minutes if you use the large and not the large Pro it's going to take you about 6 minutes and everything else is doubled in time as well but once that 3 minutes is done you'll get the log read out after about an hour the training will be done and we'll have created a folder for the model that you've created so you can see on this machine I've created the Bob D one and then I also created one from Tracy if I open this folder you're going to find the model files in various stages if I Mouse over the names you can see that the main file here Bob doyle.

saafe tensors that's the final model that's got all 1,000 steps that it defaults to and these are the models saved at various steps during the process 250 steps 500 steps 750 steps but I found that for the most part the final one is fine although I very often will download the final one and the one right before it these others you don't really need in order to use these models so just click on download the safe tensors one and then I'll download the 750 step one as well and now I have them on my local hard drive so now I can close out this machine and now I can test this L right here on mimic PC I'm going to open up an instance of a program called Forge which is already set up to run flux models and lauras and I'll leave a link to this particular workflow in the description so that you can go right to it if you'd like start this machine here's what Forge looks like now this is is an instance that I've used several times and so I already have some models loaded up in here but let me show you how to get this Laura up here so that you can use it under the file tab you have these various folders and what you're going to want is the models folder you're going to open that up and then inside the Lura folder if this is your first time running this you won't have anything in here this will be completely empty but to add lurs to this you just click this little up Arrow here and then click here to select the files I've got these lws I've created here I would just select them and click on open and then they will upload once they appear here after the upload you're going to go to the lus tab and you're going to click the refresh icon here that's going to make sure that the latest thing you've uploaded it appears to use these luras all you need to do is click on them and the tags necessary to make them work will appear up here in the text prompt and then you can just write around them for example the model that we just trained is this one here Bob Doyle so I'm just going to say Bob Doyle posing for a professional head shot in a steampunk bar I'm going to click the generation tab here so I can make sure my settings are right all of this is already set up for flux and the dev model which is what your Laura was trained on if you didn't change anything in the advanced settings so now we have a 896 X 1152 image we're going to create one of them and we're just going to click on generate and see what we get just a reminder that all of the sampling steps and everything else is already preset for you with this particular instance of Forge I wrote steampunk bark instead of bar so it'll be interesting to see what they come up with the first time you run any of these things it takes a second to load that Dev model it's pretty big progress bar means it is at work and it will start with a really high time and it'll get faster as it goes along okay so I have a picture of a dog I didn't use the keyword up here I just used I just referenced the Laur but then I didn't say Bob Doyle person which is what you need to do so we've learned a valuable lesson here with a smiley dog let's click generate again shall We There we go that's more like it all right now here's our final image it did a pretty good job now it might be a little oddl looking this is where you can play with the amount that the Laura is applied right now we have it at full strength I could play with it and maybe do 0. 9 I could do 1. 1 and we just play around with that and that's also going to depend sometimes on the prompt so it's not always like this is one siiz fits-all but it does get a very good highly detailed model this is very high res look at the skin texture on that that is kind of crazy and they got my crooked nose this did a better job than any of my attempts at my face on the SD 1.

5 model there were a few faces that I had problems with getting accurate and this was one of them but this is done a very nice job here's the Next Generation look at that hair but that's the hair that was in those pictures if you look at them again they're pretty wild but still very nice highquality image I couldn't be happier so that's how it works with a person what about a cat or an animal but in this case a cat in this case my cat particularly Lucy my orange tabby so I'm going to make a model of her I'm going to call this Lucy tabby cat I could probably do something called Lucy tabby cat but in this case I wrote Lucy tabby cat and I use the same thing for the trigger words Lucy tabby cat you want your trigger words to be something that that's the model is probably not already trained on so you want something unique and then I drag 20 images of my cat wearing the pain collar just whatever look I want her to maintain I want to make sure that most pictures have that so we've got the pink collar in every single one if I had one criticism in retrospect wouldn't choosing these cuz again I did this a long time ago is that she's very different sizes in here we have her as a kitten and we have her full grown doesn't really cause a problem but it could on down the line now in this case because it's not a person I decided to use the captioning and all you have to do to do that is click this button and it will just start its work it describes fully what's going on in the image the cat but yes where is it what's the colors around it this again helps the model be more flexible when it's time to generate images here we have more detailed descriptions an orange cat laying on top of a bed next to a pillow with a pain collar around its neck in the background there are a few objects on the table and a watermark on the image I I guess that's the time up there on the top an orange and white cat laying on top of a bed you get the idea and then it says trigger and then that's the Lucy tabby cat now we have 20 images and 20 captions I mentioned the dev model before if you wanted to train it on the other the quicker Schnell model you can do that by clicking here but generally it's agreed that the overall quality of the dev model is better so go ahead and do that even though we've selected a card with 24 GB of V Ram you still want to click low V Ram because when I tested it without doing that it actually was looking for something like 90 GB of memory or something like that anyway it didn't run all that to say for the most part you can just leave that closed and not even deal with it once you generate these captions you can just click start training and wait I don't actually have the Lucy Laura loaded onto this particular uh instance of Forge so why don't we do that now we'll click on select files we'll find Lucy tabby cat and you'll notice the progress right here Lucy tabby cat safe tensors now that that's done we can go back over here to the Lura Tab and click refresh again and now we will find Lucy tabby cat in the list actually let's clear this out then we'll click on Lucy tabby cat we'll be sure to use the trigger word now Lucy tabby cat cat Okay reading a magazine by the fire go over here to generation this seems more like a landscape image to me so I'm going to take these dimensions and flip them just by clicking this right here and click generate all right there she is that's her now I know all orange tabby cats look alike but you know we all know our babies right now because this is a flux model we can do fun stuff like adding text so how about Lucy tabby cat at the beach laying on her back in the sun next to a sign that reads pet me and generate okay maybe we won't be laying on the beach let's do a normal pose there were no samples of her laying on her backs let's put Lucy tabby cat cat at a coffee shop sniffing a steam Deming cappuccino there is a sign on the table that says no cats allow generate and you'll notice we're not using any negative prompt either and you can't even do that with these flux models you can't type anything here all right the obvious problem there is that's not her let's play with the strength of the prompt because we know that she's in there we just saw her so I'm going to just say 1. 2 there she is we didn't get the no cats allowed in this particular thing either but we got luy back now flux loves way more elaborate prompts than what I'm giving it so that's why I like the tool called glyph glyph allows you to take a simple prompt and make a really cool flux prompt out of it so let's take that lame prompt I created before I'll just say cat at this point because it won't know what the keyword is over here I'm going to leave it on hyperrealism I'm going to copy that prompt go back and type it here and paste it here I'll remove the hyperrealistic scene there and I'll say Lucy tabby cat just perched on polished mahogany blah blah blah no cats Tallow let's back her off now just take her back to normal one string because sometimes when it looks like this there's certain aspects of these renderings that it says oh the Lara is at play too much it's kind of making things a little yucky so we'll back off a little bit give more strength to the prompt and There She is again totally bored to be there normally the flux model does a great job on text I think I might need to play again with the balance of the prompt itself giving various emphasis here and there maybe if I pop this up a bit see if we give more emphasis to the text there we go no cats allowed and she's completely ignoring that and falling asleep in the cappuccino as I said you don't have to use these luras here on mimic PC you can use AI toolkit to quickly generate these luras like I'm doing and then use them on your own installation of comfy UI or Forge or automatic 1111 or anything that accepts luras let me show you so now I'm on my local machine this is one of many flux based workflows I have this is one I will probably eventually send up to mimic PC to share but I'm not quite done with it yet but it'll serve its purpose is here and there's plenty of flux workflows for you not just in Forge but also on comy UI at mimic PC to try your luras out but I just wanted to show you what's happening locally to use these luras in comy UI you're going to load them into the Laura models folder for your comfy UI installation in most cases you're going to drill down into comy UI and then models and then there'll be a folder called luras now in my case though I've got all my models loaded into a previous installation of automatic 1111 so I have them pointing to that Laura's model folder you just need to put the file wherever you're used to putting your luras same goes for if you're using these lauras on mimic PC you will upload them just how I showed you to upload them to forge if you've run a comfy UI machine on mimic you'll load those luras into that folder I've loaded the Bob Doyle Laura that I created before and I put in this prompt here and I'm just playing with getting the settings right I went to glyph and I said 60-year-old man with brown hair waiting in line for a movie it created this prompt which I pasted here I'm going to generate it again cuz I made some changes to the previous run I ran a prompt but once once again I forgot to put in the trigger word it was getting pretty close anyway but instead of 60-year-old man and I'm 58 at least for 20 more days I'll put a weathered Bob Doyle with salt and pepper brown hair all right I'll make sure the Laura is chosen in my Laura loader here right now I've got the strength of the model at 0.