ComfyUI Outpaint workflow #comfyui #outpaint #workflow

hi today I'm sharing with you a workflow for out painting for those who don't know out painting is another word for adding new pixels to an image while completing the content to match the original image we usually want the connection to be invisible and harmonious as you can see in this example we will of course start by loading the image to which we would like to add new parts I added here these nodes that allow us to see the amount of pixels lengthwise and horizontally before and after the resize I use mix laabs resize image node which allows you to keep the original proportions of the image you just need to define whether it is a portrait image or a landscape image and here you see the new resolution you can change the size according to the computing power you have from here we continue to the next group of nodes here we prepare the image for our painting first thing pad image for out paint here you choose in which direction you want to enlarge the image an important thing to pay attention to is the Feathering here you control the transition between the mask and the area without the mask if I set the Feathering to zero you can see that the transition now between the black and white part of the Mask is very sharp and this can greatly affect the final result note that at the moment the transition is very noticeable in the final result let's bring it back to 60 you can play with this number depending on the result but I wouldn't put it too low let's see what we have here to understand the role of these nodes let let's put them in bypass and press Q as you can see the result is very far from what we want to get the main problem is that the model does not have any information in this area that we added by using the fill Mast area node we can fill these pixels with information from the original image as you can see it is created by smearing the edge of the image and if we now click on Q you see that the result is completely different and much much better than what we got before it will probably be difficult for you to see but there is some very small and weak line in the connection and for that we have here the blur Mast area which will help us get a smoother and more harmonious connection between the pixels in certain images you may also want to play with the amount of blur so after we prepared the image for out paint we can enter the latent space I work with the jugrnaut lightning which is quite amazing that you can achieve such results with such a small amount of steps of course you are welcome to try with your favorite models positive prompt and negative prompt are empty in certain situations it can help to Define in the positive prompt what you are looking for but in most cases you will get a good result even without a prompt all of this goes to V and code and in paint conditioning which is part of a package called comu in paint nodes I will of course leave a link in the description to the GitHub of the project and a link to the page with the models that need to be downloaded on the left side we connect the image the mask the props and the vae and on the the other side we have two outputs of latent latent and paint connects to apply focus and paint to which we also connect the model and the patch which are actually the models we downloaded from the apply focus and paint we connect the model to differential diffusion and from there to the case sampler which is adapted to the model I chose the second latent is also connected to the case sampler in principle that's all but you can see that workflow continues a little longer let's understand why and for it to be clearer I will use another picture in this case let's say that I want to expand the image to the left side so I will add the desired amount of pixels to the left preferably numbers that are multiples of 64 and press Q so we see that in terms of the out paint the result is very good let's connect this result directly to image compare and as you can see even though these parts of the camera were not at all close to the mask they also changed a little and this probably happened in the transition to the latent space and back to the pixel space and in order to finish with the optimal quality all in all I connected the completion and the mask that we had already created with the original image before we encoded it which allows us to overcome this problem and get a quality result so I hope you learned and that we will meet in the next lessons of course you are welcome to subscribe to the channel ask questions and like if you liked and most importantly have fun bye

Channel: PixelEasel

Views: 1,939

Keywords:

Id: j20P4hAZS1Q

Length: 4min 31sec (271 seconds)

Published: Mon May 06 2024