Stable Diffusion Beginner Guide - EVERY txt2img Setting Explained!

in this video I'm going to explain every part of the stable diffusion web UI you need to know about to generate amazing AI art we won't be speaking about any extensions but I will go through all the basic settings of text to image generation and explain them in detail let's move from top to bottom now the interface you have might be slightly different from the one that I have right now but I doubt it would matter since the names of these items are still the same so let's start with this stable diffusion checkpoint what is a checkpoint think of it as a pre-trained stable diffusion model file that is trained on a specific set of images related to a specific style you got Anime checkpoints comic book checkpoints checkpoints for realistic art fantasy scenery and many more where do you download these models you can download them from a site called civet AI once you visit the site you'll find a ton of models in here you can click any one of them it'll hit the blue download button and then download it for free the file downloaded will have either a DOT ckpt extension or a DOT safe tensor X extension once this is downloaded go to its location then go to your stable diffusion web UI install directory now visit the models folder then the stable diffusion folder and paste the file you downloaded in here now to load it you can click the blue refresh icon right here and then see the drop down menu now you'll be able to see your checkpoint click on it and wait for it to load once it is loaded it's time to move a bit down and see what this is and stable diffusion the images generated based on a set of prompts the first box here is for the prompt this is where you type what you need the box below it is for negative prompts and here you can type out anything you don't want to be included in the image we usually use this negative prompt box to include keywords like worst quality low quality monochrome so that stable diffusion will always generate Quality Art now why don't you give this a shot once you have downloaded a model go and type a keyword like women on Beach and for the negative prompts write something like worst quality bad quality monochrome disfigured hand and a couple of more keywords the keywords I type will always be included in the description to generate the art hit the Big Orange generate button down below in the large blue box to the right you'll see your image slowly being generated now that you generated an image it's time to see what these boxes to your right are the first is to load the same prompts as the last generation let's say you type something hit generate and quit stable diffusion but once you come back you can click on this button and get that prompt back again the recycle bin icon is to delete your prompt completely the image icon after that is to load Laura's what are loras they're just like checkpoints but this time is centered around a specific character setting or art style and it is very small in size compared to checkpoints for example let me download this Arcane style Laura I hit download and copy and go to the stable diffusion directory go to models go to the Laura folder and paste it here make sure to paste it in the Laura folder and not the stable diffusion folder now when I click the picture icon then go to the Laura tab and then hit the blue refresh button you'll see the Laura has been added each Laura has a trigger word you need to include in your prompt and you can find it in the download page in this case it is Arcane style I write a prompt to generate Jinx then type Arcane style and then click on the lore below now you can see a special keyword was added to the prompt now I hit generate and there you go we got an image of jinx in the exact style we needed now see how I use the same negative prompts over and over again perhaps you might think it's easier to save these prompts somewhere and generate them when we need to this is where the Styles box comes into play let's say you want to save these negative prompts simply click the save button at the very corner and enter the name of the style let's just say negative now remove everything and assume you're starting from scratch now it's all about clicking the style box choosing our saved style and then clicking on this paste icon and boom there you go your time is saved generation settings alright wipe your hands in the sweat on your brows because we're diving into something Advanced well not Advanced this is pretty basic compared to other stuff in this web UI but it will still look Advanced if you're new to all this now that you know how to generate an image it is time to tweak the settings all right let's go first off what's a sampler stable diffusion usually works by first generating an image with noise on a latent space and then slowly removing the noise there's a lot of math and graphs involved here but let me simplify it all and say the way the noise is removed and the image is generated differs from sampler to sampler the most popular Samplers are the DPM ones and Euler a but you're free to try everything out some checkpoints work better with some Samplers so always read what's on the model page if you checked my previous videos you'll see I always use Euler a or DPMS to e Karis since that works best with the models I use right next to it are sampling steps remember how I said stable the fusion images are first generated with noise and then the noise is slowly removed sampling steps determine how much noise we should remove from the image step value between 25 to 50 always works best in this case a very low step count will produce a blurry image and a very high step count can actually decrease the quality here's an image generated at three steps and 90 steps and now here's an image generated at three five steps you can pretty much see the difference right higher sampling steps also eat your VGA so be careful about that below that we got three options restore faces tiling and tires fix restore faces don't do much and only ruin the phases of the image so I don't recommend using that setting at all but you can always give it a shot tiling on the other hand can generate a tiled image if I enter a basic prompt click tiled and then hit generate you can see how I get a tiled image the highest dot fix is the most important one here it upscales your image during the generation making some tiny details like eye quality a lot better to use it you can simply click the blue check box select the upscaler you want and then set the higher steps to exactly half the steps of your image steps this is very important since it saves a lot of your VGA [Applause] foreign the width and Heights I'm sure the meaning of these are obvious but there is something you should keep in mind the higher the resolution the more power stable diffusion draws from your VGA the best solution for this is to input a lower resolution and then use the highest dot fix option to generate better quality images the next option is the CFG scale CFG scale determines how much Freedom stable diffusion has over the image for example a value of 1 will give stable diffusion infinite Freedom over an image while a scale of 15 will try to make it very close to the prompt both could reduce the quality of the image so the best is to use a middle value like 7 or 9. prove this I'm going to generate an image of a girl in a beach wearing a white swimsuit with CFG scale set to one and now I'm going to generate that same image with CFG scales set to 9. see the difference right next to this setting we have batch count and batch size there's not much a difference between these settings if I set batch count to two and then hit generate or set batch size to 2 and hit generate you see I get two results the only difference is that badge size generates images parallelly side by side you can even see this in the preview of the images that are being generated well let's say you want two images with the same prompt and in that case changing the batch count is always the best below the all of that we have something called the seat what's this about every time stable diffusion generates an image we have something called a seed value this is a unique number given for each generation and we all know despite typing the same prompt stable diffusion generates different images with some differences in each generation but by copying the seed of a previously generated image pasting it into the seed box you can generate an image closer to the previous one it won't be the same as the previous one but it will be close to test it out I'm first going to generate an image of a girl in a ballroom with long black hair hair a red dress and a blue necklace wait a few seconds and now I'm going to copy and paste the seed of the generated image which you can find here right into the seed box set the batch count to four and boom it's not the same image but you can see how they closely resemble our original image hitting that extra box will give you more options to customize that variation the variation strength is what determines how close the image should be to the image with the seed so setting it to zero will make it very very close to the original image and setting it to a larger value will change the image a lot the seed resizes options allow generating of images from a given Siege at different resolutions why usually the generated images change as the resolution changes so this will prevent that from happening and that's it I'm sure we covered all the text to image settings in stable diffusion if you have any questions make sure to comment down below and I'll try my best to answer them as usual hit that like button and also subscribe for more content like this and I'll see you all in the next video [Music] foreign [Music] foreign yeah

Channel: The AI Outline

Views: 12,250

Keywords: stable diffusion, stable diffusion tutorial, stable diffusion ai, stable diffusion ai art, stable diffusion guide, stable diffusion installation, stable diffusion gui, stable diffusion install, stable diffusion browser, stable diffusion art, stable diffusion img2img, stable diffusion beta, stable diffusion online free, stable diffusion 2, how to use stable diffusion, stable diffusion video, stable diffusion prompt guide, stable diffusion models, stable diffusion 2.0

Id: l-kbCVZJk84

Length: 9min 14sec (554 seconds)

Published: Thu Jun 01 2023