Ultimate SD Upscale Stable Diffusion Tutorial In 9 Minutes (Automatic1111)

upscaling has become a popular topic with the release of various premium tools like magnific which allows you to take an image with a low amount of detail and then increase that detail and in some cases add details that weren't originally there but there is a way to achieve similar results in automatic one using an extension called ultimate SD upscale this can be built directly into your workflow I'll be breaking down what the options do and how you can use this tool to increase the detail within your images but like the video check out the patreon and let me give it to you bite-sized so what is SD ultimate upscale it's an extension used within the image to image section of the web UI and it takes an image divides it into equally sized portions called tiles and then imp paints them this is done to build additional detail on top of a portion of the image at a higher resolution you can install this extension by navigating to the web ui's extension Tab and using the available tab to load the index and search for ultimate SD upscale then pressing install and then using the apply and restart UI button in the install tab you can also install this using the install from URL option I have a tutorial Linked In the description breaking down the whole extension tab so if you want the deeper dive on that topic then check it out this extension can only be found in the image to image tab under scripts where you can select it and it will show up at the bottom of your web UI below the script drop down now that we have this extension installed why don't we break down the options within the UI so we can understand what they do and incorporate into our workflow you have a few options in terms of deciding the size of an image after the ultimate upscale is applied and this is split between using the image to image settings using a custom size of your choosing or scaling from the images original size using a multiplier you have a number of upscales you can choose from which is the model that sharpens and increases the Fidelity of an image using a variety of algorithms I would typically recommend using the same one you use for the highr fix but you may want to experiment with Alternatives you can select between three different methods for your image to be redrawn when the upscaling takes place this essentially determines the order in which the Tils are redrawn into the image linear will redraw the Tils one by one column by column and row by Row in order chess will redraw the tiles in a chatboard pattern and this reduces the chance of seeing which are the lines between the Tils which can appear in your image then the last option is none and this will disable the redrawing of each individual tile instead presenting the image at once with minimal changes as we have no tiles to inaine the ti width and height determines the size of each tile which will be used to rebuild the image you are upscaling and the height being Zero by default means that it will be the same value as the width a higher value like 768 will result result in less artifacts with quicker image generation time as you're imp painting large portions of the image a lower value like 256 will take longer as you're imp painting smaller tiles which will increase the overall number of tiles which make up the image but in exchange you will get finer details within the individual tiles of your image I'll throw up on screen a comparison between 256 tole signs on the left and 768 tole signs on the right see you can see the differences as throwing up all of the images may make things tricky to see and the 256 image took the longest so it's likely the one you'll want to see without generating it yourself when it comes to getting those really nice details which make your image look like it was upscaled with a premium tool this is likely the feature you'll want to utilize once you choose an image and find the settings that work best for you the mask blur will determine how much blur is applied to the edge of the tiles to avoid seams being visible when the image is generated as we're generating our image tole by to and this can result in visible lines where those Tils were redrawn using mask blur can help mitigate this by reducing the visibility of the lines I've set mine to zero at points in this video to highlight those seams so you can see them in action padding will determine how many pixels of the neighboring Tils will be considered when processing a tile I generated an image with a padding of 0 and 64 but only notice a minute difference between the images but it did seem to help with seams we then have options for seams fix and you won't need to use this if your image doesn't have a visible grid of lines as this will just run another redraw pass which extends the time for generating the image the descriptions for the three types of seam fixes are less than helpful but they all essentially operate in the same way by imp painting the seam itself to try and blend it in with the rest of your image while none will disable this option and is the default value you then have three options under each Den noise width and padding which operate the same as usual but impacts the seams exclusively I'll throw up some comparisons on screen starting with band pass this option made the least number of changer City image but the blur around the seam is still visible on a 0.5d noising strength so you may need to Tinker with this option both versions of the half tole offset pass gave the same results on a 0.5d noising strength but did change the way the bra looked but the seam was completely gone from our image it would be tricky to try and demonstrate all the ways this option could work as we can't use XYZ plot and the explanations aren't that helpful but in the nutshell you should play around with this option if you're struggling with the seams in your images the upscale check boox determines whether the SD ultimate upscale should utilize the upscale options and this is enabled by default as it's required to get the most out of the extension lastly the seams fix checkbox will determine whether the options under the seams fix heading are enabled and this is disabled and set to none by default now that we've gone over the settings here's a quick guide to get you started withouts scaling your images step one will be to copy the image into the imageo image tab ensuring you also carry over the generation data then navigate to resize by and select a value of around two then set your noising strength to 0.05 step two is to open the extension and set your target size type to from image to image settings then choose the same upscaler you used for the images highr fix hit the generate button and let it perform its magic if you do a side by-side comparison you should notice a higher amount of detail in your image and hopefully very minimal artifacts as part of this very quick demonstration but now I'll break down what we just did so we can understand why it's important we use the same prompts from the image generation so that stable diffusion knows what it is we're trying to redraw and it can interpret the type of details we're trying to capture which will be the same as our original image but enhanced we set our resize value to two so we can double the resolution of our image which is important for ensuring we can actually see the details we're generating but the super important thing to keep an eye on is the denoising strength the denoising strength determines the amount of change and deviation from your original image that's allowed and this needs to be set at a fairly low value to clamp down and control the number of changes which are made if you set this value too high you will see many changes to your image which completely changes the overall look composition and details and it can take some experimentation to find the perfect value for your image after that we just ensure within the extension that we're using a sensible resolution ution and using the same upscaler for consistency now that we've covered the basics there are some additional ways you can use this extension in combination with control net to get results on weaker computers with low vram I've done a tutorial on installing control net so do check it out as it will cover everything you need to know meaning we could focus this video on using this extension the steps are the same as before but this time we will enable control net dragging the same image as the one we're using for the upscale select the tower resample option and choose control net is more important you will get the same results but in theory this should function better on weaker machines which have less vram to spare you can also take your upscaled image and run it through ultimate upscale again to get even more detail ensuring you slightly decrease the den noising strength each time to build upon previous refinements this can be beneficial for achieving super high resolutions without artifacts as you're taking it step by step rather than vamping up the resolution to the maximum in one single image generation but hopefully this video has been helpful be sure to subscribe and hit that like button so others can find all the juicy information on this channel and thanks for 7K subscribers it's really appreciated this is Biz genius and I hope you enjoyed

Channel: Bitesized Genius

Views: 3,177

Keywords: ultimate sd upscale, ultimate sd upscale tutorial, ultimate sd upscale stable diffusion, stable diffusion extension, stable diffusion extensions, stable diffusion extensions best, Stable diffusion extensions install, stable diffusion extensions 추천, a1111 stable diffusion extensions, stable diffusion extentions best, stable diffusion extentions install, a1111 stable diffusion extentions, essential stable diffusion extensions, Stable diffusion automatic1111 extensions

Id: JBOjimyy7PU

Length: 9min 20sec (560 seconds)

Published: Tue Jan 30 2024