ComfyUI - A Practical Guide to Using the FaceDetailer Node for Detailing and InPainting

welcome back to my Channel today we're going to carry on from where we left off last time and see if we can improve on our image so where we left off we left off with the stacks of luras stacks of models stacks of multi upscales and we ended up with something similar to this so one of uh the viewers mentioned to add film of and uh film photograph as well as photo of these um hold up I don't know if they need to be in quotes because none of the others are in quotes I'm going to remove them from the quotes free them up and we will cue The Prompt and see what happens okay let's bring this up so we can compare I should have compared before I added this I guess we can uh remove this and run it again up scale okay so we got that one there we're going to remove this x run it and we can compare I know my audio is not that good in the videos so I'm hoping to fix this using a separate recording device namely my phone to record the audio so let's check these two out so this is without the new terms or prompts and this is with the prompts there is a slight difference I reckon with the new prompts it looks slightly better especially the eyes but I'm going to try something else we're going to use face detailer so let's add face detailer this video is about face detailer so we' got the upscale we're going to upscale first then apply the face detailer to the image oh oh actually let me try one more thing let's try one more thing before we use face deter reduce that to fre q and see what the difference in this image output is using a CFG of 3.0 instead of 5.0 and I don't know why we're using D noise when it's a empty latent image and we got this one versus that one not much difference so let's change this to one this should be one especially if it's a later image in my opinion you may disagree there is a slight difference between the two of One Versus 0.9 okay so now we've done that let's uh I'm going to change the model um sorry sampler name to DPM PP 3mi SD GPU and the scheduler to car us and let's try this this might change everything completely I can see how in the preview the way it's doing things is different once this finishes we can try that so it's uh completely different image as you can see so let's stick with this and I'm going to add uh the face detailer so if you double click on the work workspace you can get the search bar face detailer so we've got the face detailer and we're going to apply it after the upscale you don't want to apply it before the upscale because then you're upscale the details if you apply after the upscale then details will be applied to the upscaled image which I believe is a lot better so there is that save image okay now to check um what I'm going to do is we've got move this over here just to get everything read uh in order so we got Laura going to apply the model to there the clip to there so you got the model and the clip there now we got the vae and we're going to apply that from the load vae to here positive got the two prompts here got the conditioning positive to the positive conditioning negative to the negative okay so that's everything from this side added and go back here and we'll need a a bbox detector so what that is is an ultra litic detector provider and we're going to connect that to bbox we got this sgm detector we can apply that to too it's optional but I usually do not add it then you got the Sam loader Sam loader which you also can apply but you do not need and over here I'm going to move this over here and as this is an image from this the face detailer that output we want to see what the image is before the face Dy alert gets to it so we're going to apply the name to this to so we got two images here one is let's rename this uh where is it properties no not properties there should be an edit no title okay save image after upscale okay what the title title save image after face detailer title okay so we we can compare the two differences now um this this one also uses a seed there's the seed currently set to zero and uh case sampler also uses a seed we're going to use the same seed using seed everywhere so right click um convert seed to input so I check actually converts C2 widget we're going to use this now we're going to convert suit to input value there we not going to randomize fixed going to give it to this seed I'm going to convert seed to input going to give the seed to there so they both have the same seeds and we got the Primitive file name format giving both the same file name formats but they're going to have different file names okay so now we got this going to right click float right [Music] and here we can leave this as normal and run it this takes a bit longer this should populate first then that should populate with the new image so that's the first image now the face detailer is going to detail this and it's going to update one system so at the moment we've got 20 steps which is there and I'm not exactly sure what the quality is going to be like we can zoom in and double check once this goes to idle if you see the difference if you look at the eyes here and the uh lips nose eyes especially eyebrows there is a big difference eyes look really stupid here nose okay now um this was added uh the actually two days ago uh a new me method of comparing the images and there's also another one one I'm going to use an old one which is um compare image no not impact compare image see our simple image comparer okay so we're going to [Music] use where's the upscale let's move this over here upscale I'm going to use that to image [Music] one and we're going to use that to image to which goes here and we're going to do a preview image can find a okay let's add this preview image here I'm going to join these up and run this and there you go you can compare the images side by side which is quite handy but uh the other one I was mentioning which was released uh only two days ago was by um compare by RG3 so again we're going to add image B to there which is after the face detailer and after the upscale where up scale there going to add that to image a okay at first I thought I'm going a sec it's it doesn't work it's not doing anything but if you expand you can see what's going on so it seems like it's it's just doing random stuff so this is image a and if we move our Mouse across you can see image B and see the difference highlighted especially on the eyes can look at the eyes and there you go the difference this looks much more detailed obviously still needs work but if you want to compare images you can do that either this way this way or this way or you could have all three light here let's leave this over here I'm going to actually I'm going to do is add group image image comparison okay so we're going to move this here then expand so now when we move this you can move it anywhere we like and you all these will stay in the same positions unless you obviously move out so now everything in here is image comparison as you can see here you can add as many groups as you like and uh okay so we' checked that and going to up update this to 512 um going to change this to 30 steps CFG to three again just like uh um uh where's a sampler there's a sampler put them side by side and we can change the a sampler name to 3M sde GPU and that to caress everything else can stay the same and we got 30 50 steps there and we got 50 or 30 steps here so we're going to run this again I'm going to look at the comparison it's running in face detailer this is the image we need to look at this is going to stay the same because we only change to face detailer not the cas sampler so once that finishes we should be able to see what the details are especially around the eyes hopefully that does look a lot better yep I do believe that looks a lot better and you can go here and there you go see the difference now um that's face detailer uh other features we can add I found that if I got rid of the SG mm detector options and the s model so in ALR litics it's just bbox detector and and we the S we can get rid of that and we're going to run this again and see the difference zoom in here Q prompt running face detailer without uh scgm and the Sam loader you might see different results depending on what you're doing or what which checkpoint model you're using so it has improved slightly and this is a lot more photo realistic compared to this if we go here this just looks weird then you got this weird better weird better weird better if you look at the details of the eyes this is the weird one before face detailer and after face detailer okay that's face detailer uh so um at the moment we're using models luras we can put these in groups actually I think what we should do is uh have this lot in one group um so I'm going to add another group add group actually hold on let's get rid of this group so we're going to select this this this holding down control going to select all this and right click add group to select for selected nodes there you go and we can go to edit group title to model and Laura there you go so that's a different group two groups and we got empty late an image seed everywhere which can stay whatever and if we want to lock the position we can lock the position okay so we're going ulate an image what if we want to load up an image send it here and have that detailed by face detailer it doesn't matter what the prompts say so let's do this I'm going to add um a load image so we're going to add a load image and we got our image here now um we need to add a decode vae decode no v a e n code so we got pixels which joins up to image latent goes up to there and we can get rid of this latent and the Vee where's our Vee our vae can come from here oops one second need to move this up here so we can connect the the vae then we can move it back down again so we've got our vae we got image going into VA encode which is encoding the image into a latent and the latent is going into the K sampler now we're going to reduce K sampler steps because we don't want it to change the image we don't want case soundplug to change anything of the image and the noise to 0.01 we can leave everything else the same and carass I mean face detailer can stay the same and what we're going to do is run this hopefully it works so we got that before now it's going through a face detailer I don't know what face face detailer is going to do because I haven't done this before nearly there we can close this we don't need that anymore since we've got our comparison compar comparators or comparisons over here okay so let's zoom in and this is our before picture and this is our after picture somewhat different now as you can see it didn't do the whole bed or the whole face or everything on the body it's just done the face itself reason being we can figure that out by adding some previews preview image going to add a preview image to cropped refined going to add a preview image to cropped enhanced Alpha we're going to add a we're going to preview image to Cena images I'm not sure if it's going to produce anything and we're going to find out what the mask is so we need mask to mask load image mask no image Master image Master image okay so we got a mask add the preview image there and the image so now we can see the ACT ual mask let's run this take this down there so that's our image cropped refined that's the bit of the image which is the enhanced Alpha and it's going to modify all of it everything that's inside here anything outside of this box is not going to touched reason being the first four letters of this uh node face so now we can update images that we already have with a better quality face for example this like this now going to change that back to a use the prompt just get rid of the vi encode and um add what did I remove oh yeah a empty lat load latent empty here connect it up so that's what we done today thanks for watching uh hope to see you again in our next video we'll do something different

Channel: Grafting Rayman

Views: 2,500

Keywords: comfyui, stable diffusion, realistic, ai, artificial intelligence, model stacking, stacking models, comfyroll, workflow, comfyui workflow, comfyui tutorial, comfyui face detailer, facedetailer, detailed faces, inpainting, comfyui inpainting, img to img inpainting, text to img inpainting, text to image ai

Id: kbYZdMAIpH0

Length: 26min 29sec (1589 seconds)

Published: Thu Jan 04 2024