ComfyUI ControlNet OpenPose Composite workflow #comfyui #stablediffusion #controlnet

hi today we will learn about a tool that allows us to create almost any pose we will see how we use the pose we created and managed to keep perfect hands in addition we will see how you can connect parts from one image and background from another image and even remove unwanted parts from specific areas of the image let's start at the beginning basic workflow cyber realistic model positive prompt photo of a man waving in the middle of an airport latent width 512 height 768 we will connect everything of course and we get more or less what we asked for but with quite a few problems let's create more variations and it still doesn't look good enough or close to what I want so we need to help the model understand what we want to get for this we will use control net with open pose which simulates the skeleton of the human body you can use an existing image to extract a certain pose and with the help of control net with open pose copy the same pose to a new image and figure today we will introduce a website open pose ai.com that allows you to create the desired pose of course the link is in the description of the video up here choose the resolution I will choose the same ratio between the width and the height as in the image we are creating creating doubleclick with the mouse and you can change the joints to rotate and position the figure you have several images each suitable for a different type of control net and to save them simply click on the Triangular icon now that we've created a waving man we can go back to comfy now we will connect the positive and negative prompt to control net we need a suitable model open pose of course and load image to which we will upload the image we just created note that in this case you don't need the pre-processor because we are already uploading an image that the open pose understand and press Q as you can see we get four pictures with a pose very similar to the pose we have highlighted but it is still not perfect let's try to improve the result I copy the nodes of the control net and paste twice while pressing contrl plus shift plus v which preserves all the connections we will put all the control Nets together in a group that will be a little more organized and we will change the model to depth any and upload the depth map that we downloaded from the site and in the last control net I changed the model to canny and upload the image of the canny we will connect everything and press Q the result is a bit too hard for my taste and this is probably due to the fact that we are using three different control Nets with high strength of one so we will lower the strength in all of them and try again and here we need to check different strengths depending on the image and the model in this case we get three images out of the four that we can move forward with in terms of the pose but their background is not interesting enough let's create an image for the background and connect everything together we will copy the nodes we need and paste with all the connections in the positive prompt I write photo of an airport Hall [Music] let's change the seed and see what we get we'll continue with these images first of all we need to select the specific image from each batch and we do this with the image from batch note that the batch index starts from zero so the first image in the quartet is actually zero because I want the third image in the quartet I choose index 2 we also want to cut the guy from the background so we'll connect remove background great now we'll connect the guy to the background we created here too we first need to select the Right image from the quartet in our case the fourth image index 3 and to be able to combine the two parts we will bring the image composite mask node to the destination we will connect the background image to the source we will connect our guy and mask to mask and connect to the preview node excellent as you can see the combination is very noticeable and not smooth and we will fix this with a second pass back to latent space with vae and code connect to K sampler lower Den noise to 0.5 and change the seed great you can see that we keep the pose and the figure and the connection to the background is much more [Music] harmonious now we need to improve the quality of the image from the latent I connect directly to the NN latent upscale we are in model 1.5 and I choose up scale 1.5 and from there to another K sampler change the seed and significantly increase the denoise which means that we may get surprises in relation to the previous image as you can see the hand position and the figure also turned out really well and we added some figure from behind which is a bit disturbing but it's something that's pretty easy to fix and of course let's do it with compy I choose the load image node and copy our image open the mask editor and mark the area I want to clean we'll pass all this through llama remove object which will do the job for us and we manage to get rid of The Unwanted passenger here you can see several examples of characters and backgrounds made in this specific workflow found in the description of the video you are of course welcome to subscribe to the channel ask questions and like if you found this video helpful and most importantly have fun

Channel: PixelEasel

Views: 2,187

Keywords:

Id: 8Kdd3ln0gmo

Length: 6min 7sec (367 seconds)

Published: Thu Mar 14 2024