ComfyUI Image Upscaling: Unveiling the Best Methods! (ControlNet, UltimateSD, Iterative, HiresFix)

hi everyone welcome back to data leveling so in today's video we will see the different ways where we can upscale our images using confy UI what we want to achieve is like the image on the most right where it looks sharper crispier and it just looks better overall for those of you that are new to stable diffusion image upscaling how this works is that the algorithms will diffuse information across the entire image iteratively refining the details without introducing unwanted artifacts all the models and links that I used will be provided in the description I assume you have already installed confy UI and confy UI manager and if you have not you can watch my previous video on how to install those all right so let's get straight into it the first thing we want to do here is open up config UI manager and search for impact we will be installing the impact noes next we will search for ultimate and install the ultimate stable diffusion up Skiller noes lastly we will search for control and install the control net processor notes now that we have installed all the dependencies we will first generate a base image you can choose a different checkpoint model and promt here but I'll be using the Epic realism model to generate a portrait image I will also set the seat to one with fixed control so that you can reproduce the same results we will first see what a basic image upscaler does and we will Target to upscale the model by two times we can see that in the full image that there are a lot of pixelation in the image what we can try next is to upscale the image using an AI model we can drag out an upscale model loader note and select the upscale model you want to use so these models are trained on a massive amount of blur and sharp image bearing specifically for the upscaling purpose the models can be downloaded from this website over here in my case I am using the four times Ultra sharp model and because of the way this note is designed we cannot change the upscale Factor so we will have to down scale the image by 0.5 to have a fair comparison of the results another way we can upscale an image is by upscaling the latent from the case sampler and decode it into an [Music] image we can see that the result of a direct latent up scill outputs a blurry image and this is expected I will show you a better way to upscale latens we will be using an iterative latent up Skiller that can be found under impact pack upskill section this method of upscaling will provide more control and Precision in the upscaling process as it gradually increase the size of the latence rather than doing it in a single jump for the up Skiller note we'll be using the pixel K sampler upscale provider and for the parameters I'll set it to be similar as the original case sampler while adjusting only the sampler steps to be 50 and the D noise to be around 0.1 this denoise refers to how much change you are willing to accept with respect to the base image as the quality of my base image is already quite good and have most of the features I want I will want my upskilled image to be more consistent with the base image therefore setting the denoise level to be lower if your base image is of a low quality at around 128 or 256 pixels you want to set the denoise strength to a higher level of around 0.4 to 0.5 this is because your image is probably ly blurry and will require additional artificial details to each neighboring pixels of your base image as for the steps of the iterative upscaler you can set it to a number between 1 to three these steps refer to the number of upscaling iteration so in our Target upscale factor of two and our steps being two it will increase 0.5 on each iteration other than iterative latent up scaling we can also perform iterative image upscaling using this note over here the input pixels should be linked to the base image output and everything else can be kept the same we can see that this image for the latent and image upscaling using the iterative way is sharper and not as pixelated as the previous examples the next method we will be using is the ultimate stable diffusion method we can search for the note by typing ultimate I will be reusing the same parameters used for the sampler configuration and keep everything else as default we can see that the result from the ultimate stable diffusion upscaling method is very sharp and and able to keep most of the details but if in some cases where the upscaled image have too much noise and straight too far from the original image what we can use is something called a control net when you click add notes and go to the control net pre-processor section we can see a lot of different options I will use the ox pre-processor as it is kind of like a base note that lets you change your pre-processor option easily if you are new to control net how a control net works is that you will retain the image details based on the guiding image that you provide this guided image is not your base image in its raw form it has to go through some pre-processing steps to bring out finer details of the base image like edges shapes and Contours for the control net to work properly in this example I'll be using line art pre-processor and I'll will have to use a corresponding line artart control net model model this is an example of what the line art guiding image will look like as we can see the shape and edges of the model is clearly defined and these details are what we want to keep in our final output you may also test out other pre-processors and see which one is able to draw out finer details of your image the next step is to use the guided image with a control net model we can do that by creating an apply control net note the input image will be linked to the guided image the input conditioning to the positive prom and for the input control net model we will use a control net model loader note since I'm using line art pre-processor I will use a line up control net model this is the website where you can download all the control net models the models are the file that ends with. pth I would also recommend to install all the models so that you can try all kinds of different combinations we will then link up the output conditioning to the positive prompt of our interative case Samplers and the ultimate stable diffusion upscaler noes the control net strength controls how much of the guiding image is followed relative to the prom higher strength means that the output image Will Follow The Guiding image more and the lower strength means it will follow the prompt more now if let's say you do not want want to use an upscaling model or any external custom notes you can use this technique called high resolution fits what this does is that we will go through two rounds or two Paths of case Samplers after the latent up skill step the Denise here have to set to be at least 0.4 for the image to be non [Music] [Music] pixelated h personally am not a fan of this method due to the huge Den noise requirement leading to the loss of fine details by looking at the comparison of the results we can see that the iterative methods and the automate stable diffusion upscaler method works better the results are also an upgrade to the default based and model upscaling method for the high resolution fix method the result may not be as good if we are comparing the details to the base image but if we are using it with respect to the prom the image looks sharp and can be used directly I will now change the prom a little and we can see that the upskilling effect still works pretty well if let's say you already have an existing image that you want to load that is not AI generated you can simply replace the case sampler with a load image note and replace the decoding note with an encoding one I will use a low rest image of the AI goodat Andrew and try to upscale His Image to a higher resolution personally for me I think that the best methods are the ultimate stable diffusion up Skiller and the iterative image or Litt upscaler both with a control net if you want to keep things simple just go with the ultimate one the upscaling model also plays a big part in the kind of details you will have so try out different upscaling models models to see which one produce the best result for you if you learned something from this video do help to leave a like And subscribe for more content like this it will really help the channel grows and serves as motivation for me as well if you face any difficulties following the videos do also leave a comment and I will try my best to help you and remember don't stop leveling up

Channel: Data Leveling

Views: 9,218

Keywords: IterativeLatent, IterativeImage, SDUltimate, HiresFix, HighResolutionFix

Id: OnsWREg45yw

Length: 10min 54sec (654 seconds)

Published: Mon Jan 15 2024