Fooocus Tutorial Part 2 - Tips for Styles/Models/LoRAs/SDXL Turbo

in my first video on Focus I did a pretty broad overview of the software going from the install all the way to using the basics of each feature I wanted to make this video to go into more detail about certain parts of focus that might still be confusing and adding a little general information about stable diffusion itself just to help wrap your head around why and when to use certain settings I know that telling you exactly what to click and what settings to use only gets you so far and if you don't know what to do after the tutorial is over then it can still be pretty frustrating all right so start up your uh Focus if you need to install it feel free to watch my previous video to get that done and catch up on the basics of focus a link for that will be at the top now just like before you can immediately start uh typing anything you like and generate images I'm going to type futuristic Female warrior all right let's click on advanced in here we see our three speeds aspect ratios image number negative prompt and random seed check box one thing you may have noticed is there isn't a 9x6 or 16x9 ratio a reason for that is the resolutions work best when the numbers are multiples of 64 but 1344 by 768 is very close to 169 ratio and same with 768 x 1344 is close enough to 916 aspect ratio so use those if that's what you are trying to get for this I'm going to turn random seed off so the generations stay on the same seed and I'm going to put the image number to one just for test purposes now onto the Styles as we saw before there are a lot and I mean a lot among these Styles is the focus V2 style this one is different than the rest in that it uses an offline gpt2 AI model to help improve the quality of image generation I'm not certain if the other six Focus Styles have the same treatment but they do work well for getting realistic and cinematic photos all the rest of the Styles here are basically embeds of very well written positive and negative prompts that are added on top of your main prompt a user by the name of sansu actually compiled a list of picture examples of the positive and negative prompts that come with each of these Styles the link is in the description if you would like to check it out additionally someone else by the name of aat created a compilation of their own using pumpkins I like this one as it really gives you a nice visual of each style not every style is included as this was created a few months back but still has around 184 of the more than 280 available the link for this is also in the description you can mix and match these Styles if you add a style and it doesn't seem like it's taking effect try typing that style in your main prompt and you should get something closer to it if not enough remove some of the other styles you can really get some interesting and fun combinations here moving on to the model tab we got the base model also called a checkpoint these are the main components of stable diffusion they are models designed for generating images based on what data it was trained on for example if the training data doesn't include images of cats the model will be unable to generate a cat image the same goes that if you only train a model to generate cat images then it will only generate cats hopefully that makes sense Juggernaut is a good overall model that is geared toward photo realism and cinematic images dream shaper is another good model it can do photography and realism but is also trained on anime and more artistic and fantasy Styles either one of these can do a variety of other image types but will have its own strengths and weaknesses as mentioned before Focus can only use sdxl based models sdxl is a newer more powerful version of stable diffusion but SD 1.5 tends to be more popular because of how much how much more time it has had to mature and be optimized by the community there are many more tools and Extras available for it right now but eventually sdxl will catch up I will show you how to add a new model if you wish let's go to civit AI you can download and use any sdxl model here that you want want to give a try to find an sdxl model go here and change this filter setting to sdxl 1.0 then click checkpoint this will filter out everything else you can grab whatever you like but I am going to select dynavision XL you want to make sure it says the correct base model sdxl 1.0 up here you will see past models and sometimes there are models for other versions of SD for now we will stick with this one sdxl models are usually pretty large so 6 GB isn't a surprise I will download this to the focus stable diffusion model file go to your focus files and then go to models and checkpoints this is where our main sdxl models will go also a note when sdxl first came out it was generally used along with a refiner the refiner acted as something that would put the finishing touches on your image the checkpoint would take it most away and the refiner would fix the last bits of detail But as time went on sdxl models became more fine-tuned and the use of the refiner for most of them is becoming obsolete when you download an sdxl model check the details down below you can see here that this checkpoint is designed to not use the refiner before leaving this website I want to get one other thing allora let's go back to the main page and change the filter from checkpoint to Laura still leaving sdxl 1.0 checked a quick overview when talking about models most people will be talking about the checkpoint but in reality all the extra components for SD are referred to as models and when talking about base model they are referring to the version of stable diffusion you are using in our case our base model is sdxl 1.0 and our model checkpoint that Focus comes with by default is the Juggernaut checkpoint and now we are going to look at Laura models all confusing I know so what is Aura Aura can be thought of as a patch that you add temporarily to the checkpoint you are using it will influence the outcome depending on what the Laura is designed to do some of them change the art style this one can control the the amount of detail in an image this middle image represents the unaltered image once you have the Lowa installed you can change the setting to remove detail by putting the strength or weight of the Lowa into the negatives or add detail by having the strength into the positives for Focus the way we change the strength will be with the sliders more convenient there are also luras designed around a specific celebrity or character this one is designed to give you Sarah Miller from The Last of Us we will try this one download the model safe tensor this one will go into your focus model Laura folder also if you scroll down you will see some helpful tips to using this Laura using the suggested prompts here will give us the best results okay we can get back to focus now and try these out when the downloads are finished we can come over here and hit refresh all files and they should now show up in the list before anything else I want to mention the Laura that is on by default sdxl offset example Laura this in addition to the refiner was often used with sdxl to get a good image it isn't necessary to use but more of a personal preference so what does it do the example Laura adds offset noise to the model when used it will extend the images contrast range of brightness to Darkness it can be helpful in Darker images to get deeper blacks in my own experience of using it my results are random I generally leave it off okay so the refiner as said it isn't necessary most of the time but I can show you quick what results I get the official refiner is not included with Focus if you do want it to try for yourself you can get it here there is a link in the description for this too click files and versions and download this one put it with your other checkpoints okay I'm going to start with this image then add the refiner in the text you can see the refiner is suggested at 0.8 strength so that's where we will put it and generate so this is with the refiner this is without with without so there is a change but I can't say I see anything that makes me feel this is a necessary step as it adds another process and more time between Generations but it's up to you okay removing that for now let's try Aura let's add the sahura this slider next to it is the weight or strength this is how much the Laura will influence the picture at one we are at 100% but you can go all the way up to two or as low as -2 but don't do that or you will get crazy results usually when using Aur you will stick to the 0.1 to one range so leaving the Laura at one you see we get exactly what we are looking for a picture of Sarah not too bad if you are not getting good results remember that the Styles you have chosen also influence the image you can also use the suggested prompt from the Laura main page not bad uh we also have our new checkpoint load that and try seeing what happens the first time we change a checkpoint the generation will be slower but after that it will speed up again wow that's a nice style the D Vision 3D Style with the Sarah Laura I will remove the Laura and see what we get nice so with this checkpoint you will get a style more like this okay onto the advanced tab the guidance scale can be used to give a more artistic outcome what you will generally get is something that will have deeper colors you will often see people call this burned or overbaked because the colors will start to get cartoony and if you are going for realism then you won't want to have this set too high sharpness again pretty self-explanatory here's some examples subtle changes but you can see it in the armor and the background okay last time I skipped the debug I'm not going to talk about it too much here since that would make the video extremely long but I want to give one tip for those who want to try out sdxl turbo sdxl Turbo is the newest thing and it basically makes image generation much faster I have the dream shaper sdxl turbo model already but you can go back to civid Ai and change the filter from sdxl 1.0 to sdxl Turbo and choose a model these go with all your other checkpoints so let's change our checkpoint to the new sdxl turbo model now go to advance Tab and change guidance to 1.4 and sharpness to three check the debug mode we need to change two settings here go down to sampler and change this to dpmp sde then change Forest overwrite of sampling step to six these are the settings to get turbo mode working and start generating the first time will be slow as it's loading the model but the next ones will be much faster this will be able to create images in only six steps steps are basically how many iterations the AI runs through the image each time it gets closer and closer to an image other models will take 20 to 60 steps to get a decent image using turbo is nice for slower gpus these images may not be as detailed as regular models with more steps but they still can produce excellent results for so little time we do have this whole other section down here called input image I did go over the basics in my first video but I will eventually do an in-depth video about it for now I going to leave this video here I hope for anyone who made it this far that you learned something valuable see you in the next one thanks

Channel: Jump Into AI

Views: 9,160

Keywords: fooocus tutorial, stable diffusion, fooocus tutorial for beginners, stable diffusion tutorial, stable diffusion for beginners

Id: OXKvMTGVzJY

Length: 11min 48sec (708 seconds)

Published: Sat Dec 23 2023