ComfyUI Tutorial: Consistent Character In ComfyUI Part 1

in this video I'll demonstrate how to create realistic characters with consistent face inside comy UI let's start building the nodes from scratch the first step is to load the checkpoint my preferred checkpoint for generating realistic images is photopia XL which is available on civit ai.com currently I'm using their Type 4 version the version 4.5 is also currently available let's proceed by loading the IP adapter and selecting the clip Vision [Music] model please note that you might have a different name for this model in your setup next we load the IP adapter model I prefer using the IP adapter plus s xdl model we can now load two images using the batch image node which will combine the two images [Music] however this method doesn't allow me to apply different weights to the images therefore I'll use a different IP adapter node for that I'll utilize the encode IP adapter image node to obtain the embeddings with custom weights for each image the first image will be directed to the first embedding the second image to the second embedding and so on however clip Vision tends to crop the image so I won't crop these directly instead I'll use a prep image for clip Vision [Music] node to duplicate this node for the second image simply hold down the alt K and click for interpolation I'm selecting Lanos we have various options for selecting the image positions in this case I'll choose top because the face is positioned at the top of the image our goal is to create a unique face by blending these two images which will eventually become our consistent image connect to the clip Vision next we'll require a k sampler connect the model and introduce an empty latent image since we're using the sdxl model here set a higher resolution also set a batch size of two for now okay let's write the text [Music] prompt connect vae to code connect vae to the checkpoint vae here we will get our image but I'm going to use a face restore node here to refine the face at this point we will generate our image however the quality of the face might not be satisfactory as we'll see in the preview to refine the face I'm going to use a face restore node [Music] [Music] let's set the weights here we can be creative and test out different weights let's try [Music] something [Music] so here's our result and here's the face restored version let's compare it with the reference image used and see if it matches with any of these images [Music] so the image is a complete blend and a complete no person now let's compare the generated image with the reference images used and see if it matches any of them as you'll notice the resulting image is a complete blend and doesn't resemble any specific person we can attempt this process multiple times until we're satisfied with the result if we like the face that's been generated we can use it to create consistent face images [Music] [Music] to do this let's save the image to our local folder so let's begin to create unique and consistent images using this face for this we will use reactor face swap node which has a built-in face restoration so let's get rid of this face restoration and add a reactive face swap [Music] [Music] node let's go to pixa Bay to to find another free image I've loaded the image [Music] here okay all set let's run [Music] this all right here's our image it seems there's an issue with the eyes but we'll address that in subsequent attempts first let's compare this image with the reference [Music] images it appears to have Incorporated details from both images such as the hair from one image and the red and white colors from the other now let's figure out what happened with the eyes I suspect there might be an issue with the face model we created earlier to rectify this let's first create a new nobody face for this I'll disconnect the reactor node and reconnect the face restore [Music] node we've now generated a new face let's save this face and we'll use it as our consistent face in the reactor node all right I'll conclude this video here stay tuned for the second part where we'll delve even deeper thank you for watching and I trust you found value in this content

Channel: TECHZUHAIB

Views: 1,362

Keywords: COMFYUI, IPADAPTER, STABLEDIFFUSION, SDXL

Id: GIo7INXaPAw

Length: 7min 48sec (468 seconds)

Published: Mon Dec 25 2023