The Ultimate Guide to A1111 Stable Diffusion Techniques

the thing with the techniques is that they are not obvious but you will not believe the impact they make today my friends I will guide you through a five-step journey to crafting 4K or even 8K visual masterpieces we'll uncover the deuce and don'ts and I will share some invaluable tips and insights for checkpoint we are downloading real cartoon realistic this is one of the best models here on Civ AI for semi-realistic images Al so we will use this fantasy style a which will Infuse our images with mesmerizing fantasy effects last but certainly not least we're enhancing our images with this detail Aura a tool designed to significantly boost detail richness returning to automatic 1111 I came prepared closeup image of a female Druid in leather armor sitting on a rock casting a nature spell smiling I also included zoraa with a strength of 0.8 we start with the maximum resolution of stable diffusion 1.5 in 768 by 768 but hey Chris why not jumping directly to a 6x9 resolution like 768 by 432 this is tempting of course however low resolution sacrifice detail you will miss later on I will soon reveal a far superior approach set your sampling steps to 35 your sampling to DPM Plus+ to M caras and you batch count to eight images because we want to have a nice selection also don't use hus fix this is crucial I cannot emphasize it enough later in the video you will learn how professionals upscale now press contrl enter to render let's see what we got I like this image already this is interesting especially with the horns wow this image with a portal is great love it I like the Hound here in the background very very much and this here is just magnificent there's some real crazy nature spell going on these are all great but for the sake of demonstration this image here will work the best what I do now doesn't make any sense at first but trust me it will I save my image to disk by clicking the save icon and then click on it to download we sent our image over to the image to image tab by clicking this button this is a promising start but so far we've just scratched the surface for this next step we need control net to be precise we need a control net inpainting model you can find it at this URL just Ure to download both the yaml file and the pth file and while you addit grabs the tile model too you will see this game changer in a later Step In Action if you have no idea what I'm talking at the moment watch this video first and then come back after WS our P Druid is missing an arm as you can see and we are going to fix this now with in painting if you haven't used control net in painting prepare to be amazed hit the inpaint button and give it a moment to load then select the appropriate brush size and paint over the area you wish to alter here's where it gets exciting unlike the usual method of changing the inpaint area to only mask we will let control net take the rins no need to adjust you don't even need to change The Prompt make sure your sampling method and steps are the same as before prepare for some variation by setting our batch C to four we will leave the D noising strength untouched for now dive into the control net dropdown activate it and filter by inpaint select either inant Global harmonious or inpaint only plus llama for the pre-processor each of them does wonders in its way ready let's render Behold The Magic of of control net we will take this image here although the hand isn't perfect I've got another trick up for that in my sleeve I will teach it to you when we are progressing our journey into upscale territory soon now send the image back to image to image you can sure fix a lot of things with in painting but one thing apart from hands where stable diffusion stumbles upon is when it comes to text here is where this week's sponsor shines Storia lab by Storia there are two box is impressive but it's a textify tool that truly captures my Fascination and here's why you can fix any spelling mistake made by AI image Generation all while preserving the original art style simply upload your image create a text box over the area in need and type in the correct text once you hit apply the AI Springs into action generating multiple version of the corrected image isn't that impressive signing up is effortless and they even welcome you with free credits Storia also boasts an impressive cleanup tool designed to seamlessly remove any undesired elements from an image taking this Bioshock inspired image as our canvas we simply highlight The Unwanted figures using a brush signaling the AI what to erase hit apply the result is a remarkably cleaned up image as for story our pricing it strikes the balance between affordability and unlimited creativity consider the immense value this brings to your workflow especially when collaborating with clients on projects I cut you a sweet deal of 10% of your existing subscription for the first 6 month just write a mail to Founders story. thanks again to Storia for sponsoring this part of the video now prepare to be amazed as we Elevate our work from its current state to something extraordinary we are boosting our resolution to 1,368 by 768 to achieve a 60 by9 aspect ratio set the D noising strength to .9 and yes you heard that correctly the rest of the settings can stay the same here's where the true magic unfolds activate our unit and check the upload independent control image option we will upload the image we saved earlier right here now select inant but listen closely choose inant only plus llama don't choose Global harmonious this time the later would alter our base image which we do not want and sure control net weight is set to one with the control mode set to control net is more important set the resize mode to resize and fill failing to do so could lead to strange images let's it render in certain instances removing the prompt might be beneficial I suggest trying it with the prompt initially out of all these images this one here stands out the most don't hesitate to further experiment with this it's time to take our resolution to the next level we need to switch the TP to a resize bu here you will adjust the scale between 1.5 and two depending on the capabilities of your graphics card personally I opt for a setting of two on my 4080 we still keep the D noising strength to 0.9 and yes this is still correct in our control net tab uncheck the upload independent control image option and this time we need inan Global harmonious instead of llama experiment with a weight between 3 and 6 and set the control mode to balanced should the image details still not meet your expectations consider increasing the D noising strength to one and if that doesn't suffice reduce the control weight further to three or even lower but be vary of going below 0.25 in my experience dropping beneath the threshold should can lead to severely disorted images let's it render the details in the image already look great it even fixed our hand problem but wait until you see the last step because so far you have seen nothing remember the ti model we downloaded earlier well now it's time to use it but first a quick detour to ensure we've got all the necessary Tools in place because we missed some for the next step it's important that you turn off restore faces it's a feature that is hidden by default in newer versions of automatic 1111 so just go to settings and here type quick this should filter to the Quick Settings list here type face and select face restoration hit apply settings and reload UI now this checkbox up here should appear this doesn't make any sense now but I will explain in a moment why you need to uncheck this for our next step head over to extension tab click on available and load from then type ultimate and here install the ultimate SD upscale extension I can't wait to demonstrate this incredible upscale script to you only thing we need now is to download the 4X Ultra Shar upscaler go to this URL after downloading it you put it in your stable diffusion web UI folder under models and here under ESR again and there you put it in now are you ready for this mindblowing last step it's finally time to decrease your denoising strength to 0.3 or even lower this is trient Arrow and dependent on your checkpoint and lowers with your enabl control net unit click on ti/ blur pre-processor should say tile resample and for checkpoint it should say control net v11 tile make sure the weight is set to one and that control net is more important is set now we put our freshly installed upscale script to use go down here and select from the script the ultimate SD upscale don't confuse it with the SD upscale set the target size type to scale from image size and the scale to two times below under upscaler select the 4X Ultra Shar we just downloaded for the tile widths you should go as high as your graphics card is able to manage I usually go with 768 but why do we do that because we do what's called a tile up scale so we want as little tiles as possible because it means less seams which gives a clearer image in general that is also the reason we turned the restore faces off because otherwise you will end up with images like this instead of that now it's time to be amazed let's render this could take quite some time take a moment to appreciate this masterpiece truly Splendid isn't it the intricacies the depth it's clear we've outdone ourselves in crafting this gem yet if you're looking to take your workflow to even greater Heights I highly recommend checking out our next video

Channel: AIKnowledge2Go

Views: 20,476

Keywords: AI News, Artificial Intelligence, AI Tools, AIKnowledge2Go, AI Updates, AINews, AiTools, stable diffusion, stable diffusion tutorial, controlnet stable diffusion, stable diffusion upscale, comfyui, prompt engineering, ai art, upscale stable diffusion, automatic1111, stable diffusion inpaint, inpaint, stable diffusion controlnet, controlnet, stable diffusion img2img, civitai, stable diffusion ai, controlnet tutorial, stable diffusion lora, img2img, stable diffusion inpainting, tutorial

Id: wyDRHRuHbAU

Length: 11min 19sec (679 seconds)

Published: Sun Mar 10 2024