ComfyUI 07 Upscaling Latent and/or Image

images generated with stable diffusion are in the neighborhood of 500 pixels for the older models or with the new XL models they are around 1000 pixels in general you would like to have your images to be much larger on a 4k monitor we want at least a 4K image upscaling is our friend and the method I personally use most often is to do a latent upscale by two and followed by an image upscale by two there are different options though and I created a workflow where I can mix and match and see what works best for a particular picture let's have a look this is my upscaling workflow it's built with efficiency notes which were the subject of a previous video so it works from left to right here at the left we have our efficiency loader then it's followed by the efficiency simpler and then yeah we have a small image which we like to upscale so we now need an upscaler the first upscaler here is the latent upscale a right click to find it add node and then we go to latent and then we go to upscale buy or you can choose upscale latent then you have to fill in the numbers I prefer upscale buy this is the one and all I did is give it a red color because I use it also as a switch when I hit Ctrl M on the keyboard it is Switched Off and it doesn't work in this workflow I can choose to use the output image of the sampler to go to an image upscale or further down the road but normally I prefer to use latent upscale first so I switch it on we will see the images or the differences between the images in a second let me first go on explaining what we have here on the right right so we now have a latent up skilled model which is sampled and beware the simpler denoise around 0.5 seems to be the best value maybe a little bit more between 5 and 0.6 gives usually the best results so that image that we now have we send that to an image upscaler which I have over here and which I can switch on and there it goes that image upscaler we can find it over here add node and go to image and then upscaling of course and there we have image upscale by model using a model that's the one and then of course we have to load a model that is this one over here so to find that add node and then loaders and there we have load an upscale model that's the one uh that upscale model that of course goes here in the upscaler itself the image goes in there and then an upscaled image comes out what I do with that upscaled image the upscale models they often are two times or four times if it is four times now my image is already too large so I don't scale the image and the node used here is actually I changed the title it isn't it is called an upscale node so that can be found here add note image and then upscaling and this is the upscale image that I use I do now use the image one which means I can enter the values you could also choose ebsco image by then you have effector to enter but I want to use the numbers because I have a 4 okay monitor I put it on 2160 for the height and then everything goes fine for me so now I have an latent upscale and followed by an image upscale let's have a look at how that performs oh wait I forgot something we use for the image upscale we used upscale by model which means that you need an upscale model to use it and to find obscale models there's this very nice website called openmodel database just go there they have upscanners for every type of image that you can imagine and there are plenty of them so make a choice and then download that model and put it in your comfy UI model upscale folder this is our base image it's 512 pictures pixels stretched on my screen this is a latent up skilled twice note how the image changes that's because we have that D noise factor of 0.5 which means noise is added and it is again sampled so there is a change from the original image to the latent upscaled image but the quality picture quality itself is much better look at your necklace that also changed and but now we are going to add the two times image upskill after the latent upscale and now and now it is subtle but there is a clear difference maybe it doesn't show on the HD video here on YouTube on a 4k monitor I can assure there's a difference look at her necklace this is latent only and this is image and back and forth there is a clear difference on a 4k monitor okay so latent upscale followed by image upscale it works it is not all we can do with this workflow we can also do I suppose just an image upscale by four uh I can simply switch this one on then I drag this image output to this image input and now I have an image upscale by four we will see in a second what the image looks like but also we could do it differently do an image upscale by 2 which means that I down scale it again I don't scale it to 1024 and now so that sounds strange but we first upscale the image and then we down scale it again we shrink it and then the output of this one we send to the next image upscaler and I switch it on and that one I put on my monitor size oh that is outside of the view well and that is 2160. so I now have a base image generated here and I have an upscaler here and followed by a shrink and then another upscaler maybe that gives another result a better or worse I don't know yet we will have a look and also what we can do even is not send this image this image to the second upscaler we can send this image to a vae encoder then do another sampling step this time not with 0.5 because that will change the image again a lot but a smaller value of 0.3 or even less is fine and then we have another sampling step which can improve the image add a little detail and what have you not and then that output is what we sent to the final upscaler uh this workflow has all the options so let's have a look at a couple of the output images okay as a reference this is the base image again now we do not add a latent upscale we do an image upscale first by two and note the higher detail but it's also quite coarse noisy a bit image four times in one go and well a subtle difference and now image two times two that makes it smoother and then let's have a look four to mile 2 and 4 times and two times two that is a smoother image and now comes uh the final one first image upscale times two then do a sampling step and then again image times two and yeah that changes the image a little bit does it give a better quality about in some areas at this it is less smooth but in some areas it still is a bit smooth it's a bit of a mixed picture well you can just experiment the workflow is available I will put a link in the description and then you can experiment whatever you want latent or image upscale or any mix possible that's what it is about well thank you for watching and in the meantime until the next video have fun

Channel: Rudy's Hobby Channel

Views: 7,231

Keywords: stable diffusion, comfyUI, upscaling, upscale, upscaler, AI

Id: nK5oQ3fpO8Y

Length: 10min 9sec (609 seconds)

Published: Sun Aug 27 2023