How To Install ComfyUI InstantID Run With Stable Diffusion (Tutorial Guide)

hello guys so in today's videos we are going to talk about this instant ID now they have released the GitHub page and also their research paper in github.com now we will take a look at their stuff here and their demo datas also as well we will see scroll up a little bit we see the author of here is the same people as photomakers now I did the photo makers previous videos you guys can check it out and also you scroll down in here you see the multiple ID and the multiple stylus that is the similar techniques as the photo makers that using one photos and create the face with other different styles now as you can see here they are using the same set of testing data but the one unique from the photomakers is that they're using Jackie Chan and then Emma Wilson to create some other styles that is completely different with photo makers but then well when I see these two research projects I would say they are very similar but then their technique is a little bit different which the instant ID when we hop over to the hugging phase model page as we can see well in here they have also introduced their demo data and things as well but then their techniques in instant IDs they have the control net to create the face similarity and also applying the IP adapter algorithm to applying for the face ID identity for your generate image and also from here as we can see mainly we are going to download the control net models files and also the IP adapter bin files that is the two main AI models files that we are going to download so hop over to the files and versions here you see this 1.69 gigs IP adapter and then this diffusions model save 10 Source 2.5 gigs files that is for the control net so they are also available for you to test the demo publicly in hugging phase I will share this link as well but then this page is kind of busy sometimes it's just create errors even you upload things correctly so we are going to test it in comfy UI and then comy UI recently I have saw they have native instant ID support for comy UI this is made by the developer of Ip adapter which I have conf Ence of using this without any cutting into any error with the coding and looking at this one they have the basic workflow for us to test for the instant ID using text to image and in this videos we are going to test this workflow as well so first of all we will go through how to install this instant ID and right now before we do any of these installations as we saw in the message here we have to update our comfy UI to the latest versions as well so let's go over to comy UI and we will start updating our files and also this comy UI system as well so open up the confy UI folders and then we are going to execute the bat files for comfy UI we will start up the command prompts here and then wait for the loading finish you will see the comfy UI web UI system and we are going to head over to clean that up first so first of all we will head over to update the latest versions of Ki before we install the instant ID so head over to comy UI manager and just simply click updating all and we're showing all the checking of the update and installations pack for your custom nodes and other installation files for your comfy UI system and right here you see the command prompts everything is checking if there's any new versions update it will help you to do that and also your comfy UI system are going to update us well so we've got to wait for this to complete and we will start on the next steps so after that we will see the restart button showing up as we saw this is finished and we will go to comfy UI web UI and click restart button and wait for another boot up of the newer versions of comy UI as you can see it's loading and some new stuff are installing into the comfy UI system right now we are going to wait for the installations and then looks like everything are going to finish without any error and we see the blank fresh comfy UI pop up so next thing we are going to download the instant ID now they have said that the custom node have not been indexed into comfy UI manager yet so let's check out if right now they have indexed or not so hop over to install comfy UI oh there you have it they have updated the comfy UI manager custom note list already by today of my recording so so we are able to download this very simple and easily by clicking the config UI manager and just click the install button in here well if we cannot find this then we have to download the save file or GitHub project manually and located in the custom note folder but right now we can just click it the installations in here and wait for the restart right now so after the config UI install and restart we have to go over to download the hugging face to get the inside face files for the face detections there's two links here one is their Google Drive zip files it's very straightforward easy all you got to do is just download the zip files and extract into the folders that it mentions and the other one is the hugging phas so I'm going to use the Google zip files just for better easier to download the zip file and extract into this folder path so the folder path are going to be in hugging face I mean comfy UI models and then inside face so there's inside face already because I have downloaded the previous one so in the inside phase models of folders We have to create this folders the anti-loop 2 folders so let's create this one and we are going to put the zip files into this files locations so put that in this folder path and save it and wait for the download finish and we are going to extract the zp files and everything is ready for the inside face for the instant ID face detections so let's extract the files here actually we can just drag those files from here it's way easier so highlight all these files that we need and drag those on these folders and in these folders We have all set and ready and I can delete the zip files because things are duplicated and I don't need those anymore then we have the inside face third party libraries for face detection to downloads ready and also you can download in hugging face that is the same files that we have just download but the first link with Google Drive is easier just download the SEF files and you extract those files in the folder so follow this method and it is so easy to complete everything and the next step is you have to download the hugging phase of the control net and IP adapter models of insight phase and again this is quite easy but then one thing remember to do is rename those models files with corresponding functions or features of the file name so that you have easy to recognize in the future that you don't see everything is IP adapter and all is about control net now you have to create this instant ID folder and then put this IP adapter files into this instant ID folder in the model subfolder so what I do I will put instant ID in front of this IP adapter's file name so I will be easier to remember and recognize when I'm using that in the workflow so you can name other file names as you want but for me it's easier to recognize with this way and then the other second thing we have to download is the control net so Hubble foot to control net folders in the models file and then we are going to download this control net especially for instant ID I'm going to copy the name again and put that name in front of this safe tensor file files again and save that into the control net subfolder then click save everything is going to be ready now all we got to do is just wait for the download and once everything is complete then we can load up the basic workflow of instant ID to test the instant ID how it going to be running and the next step we have to remember one thing is that the instant ID is trained with sdxl so it is only supporting sdxl right now for instant ID and I'm not sure if that is supporting SD 1.5 in the future or not don't ask me if you want to know you got to ask the author of instant ID so you can go hop over to the hugging face or the GitHub page to ask for support of that but then in here they have highlight also remember at this moment this is only for sdxl so remember that when you choose the checkpoint models and your IP adapter if you using it or the control net to using with the work inside the workflow that remember select the sdxl type and also they have support sdxl turbo which means you can load this thing fast but one thing they said in here they mentioned that the training data has a lot of watermarks so maybe they are using some stock photos in the stock Marketplace and then use those to train data so use other dimensions rather than 1024x 1024 they said using other dimension then you will have no watermark on your result so hover over back to the hugging the comfy UI back to the comfy UI interface and we will see the new workflow let's say that first we are going to test the stxl turbo and see how fast it will generate with the instant ID so this is the basic basic workflow from text to image that I grabbed from here that they have available for download as well so you guys can download this workflow and test it by yourself but before testing you have to configure the checkpoint models and all other files name that is corresponding to your system so in my system here I have other name for my sdxl Turbo in here so I select my folder path for sdxl and then I have to select my face image of course I have my own AI models generate face right here so you have your own and you can use it as well so instant files here one of the tricky thing is that you don't see any updates of your name your download file name you click the refresh button it will show up your just downloaded iPad after bin files that you you downloaded so remember refresh the click the refresh button and you will have the download of your files so the same thing as photomakers they have the face analyst for you to select whether you want to use CPU or GPU using this instant face analyst notes here so connect this to the inside face point the Green Dot here and then the image features are the reference face image where we using the face and then the loading checkpoint connecting the purple dot to the models and finally the conditions positive negative connecting to our text prompt and then below that we have aend mass so we don't have a mass segmentation at this moment so we leave that as blank and then use the load control net models now remember to select the instant ID control net models that we just download so in here I have mine right here and then connect the phase keypoint processor this is the pre-processors of the instant ID because the instant ID are connecting like a control net Manor to generate a new image with the reference face so they are learning from the reference image of the phase apply as a control net and then generate passing to the K sampler and generate your new image here in the control net apply control net we are passing the pre-processors of the instant ID and then we go to the case sampler now the case sampler we have to set their requirement let's see I remember they have mentioned it's important to lower the CFG at least four or five or you can use another CFG so let's set it as 4 CFG 4 in here and click generate and see if there's any error and hopefully there is no error and initial starting we will load a little bit longer for the checkpoint models loading up and we are seeing the inside phase load and the case sampler right there so it looks like two 2 seconds in the sampling steps and there you have it we have the watercolor styles of the phase that we have and one thing I have recognized I mean I've realized is that this control net are also simulating your Source image post as well and you can use other poles in your text prompt or also the face key points pre-processor image you can connect other image in the pre-processors this part for other image post in your new image so if you keep using the same load image it will be generate the same Styles and as well as if you are changing your text prompt you will have different styles as well for art painting or watercolors into realism Styles and even if you want to change other styles you can use other checkpoints to create another styles of image in here now this is so far I have checked and investigated by changing the positions of the pose of the new generate image characters and also the styles of it you can use other checkpoint models and text prom and then using other case Samplers random seat numbers you can see there will be other styles coming out in your AI image as well and then right here I also realized that after you change the text prom to whatever in here and let's try with other sdxl models let's try the real V turbo right here I have downloaded real V version 3 Turbo and we can see other styles that is going to be generated with this checkpoint models so as you can see this similar position like a half face like the image head is like a closeup and then the image style is different because of changing the checkpoint models and we have realism Styles right now and then you can change the apply control net strength numbers to lower down the strength or put it higher strength and it will comes out with other result the lower the strength is and that will be all depends on the text prompts or your input prompts for the result image and the higher of the strength is and that will be more influenced by the reference image for the face and that will be so far what I seen in this inside phase instant ID new custom nodes and this is very similar concept as other control net models so let's say I put 0.4 here you will see very similar as the source reference image even like a close-up shot of this the pose of this is very very similar but if you lower down this it will follow the text prompt styles for the character so let's say I put this in the middle like 0.5 it will be very similar like the reference image like a half body closeup shot and then the face is right on the middle like this even the post is very similar except there's some little difference on the hand and a little creativities by the AI so you can twist this and the text prom and then you will see other result as well so let's go to using other text prom in here and you will see other styles coming up from the result so here I got the characters carrying a cat I mean two cat in here so let's tune it down to 0.4 it will be more more creative for more influenced by the text prompt rather than have influenced by reference image let's say I have like a happy face for the characters and let's see what will be happened so let's tune it down to 0.4 it will be more creative for more influenced by the text prompt rather than have influenced by the reference image let's say I have like a happy face for the characters and let's see what will be happened let's boot up to 1.5 happy face and let's see so yeah there you go a little bit smile happy lady holding the cat and let's say the smile face rather than happy that will be more face expression from this yes the facial expression is more with other text prompt you have to try and test with different styles and there you have it this is instant ID a very basic workflow that we can test around with the instant ID Technologies and how it is different from reactor phase Swap and also the photo makers and this one so far I see is based on the control net is very flexible but for convenience I will say reactor is still the most convenient the most easy and simple way to do face swap but it is up to you if you prefer using control net as the face swap then you can use that as well that is your freedom hope this video is inspiring you to try out instant ID so I will see you guys in the next videos have a good day see you

Channel: Future Thinker @Benji

Views: 8,902

Keywords: stable diffusion, stable diffusion tutorial, comfyui, stable diffusion ai, stable diffusion controlnet, instant ID, GitHub page, research paper, demo data, photo makers, face styles, testing data, control net, AI model files, installation process, update comfy UI, AI model extraction, SDXL compatibility, SD 1.5 compatibility, text to image, ipadapter stable diffusion, instantid comfyui

Id: PYqaFRLdoy4

Length: 22min 50sec (1370 seconds)

Published: Tue Feb 13 2024