ComfyUI: Tipps für jedermann | Stable Diffusion | Deutsch | Englische Untertitel

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
Hello and welcome to this video, in which I would like to exchange some life time against knowledge. In this video, I don't go into any special techniques, but just a little bit of experience that I have now collected due to your YouTube comments and also feedback in the discord or questions in the discord. There are always new people coming in again, the topic of ComfyUI and StableDiffusion, and I thought it might be a good idea to create a little update video on how to handle ComfyUI. I will now give you a few collected tips and tricks that describe the handling of ComfyUI and maybe make it easier for one or the other to handle it. First of all, we have the shortcuts. ComfyUI comes with a whole range of shortcuts and I would like to show them to you briefly. First of all, you probably already know that, but I'll just go through the list. I've sorted it out a bit so that it's a bit more suitable for the video process. But the first shortcut, which you can also see on the ComfyUI Github page, is double-click with the mouse. Then you get a search window and here you can search for terms, search for nodes that you need and then select them very conveniently. Of course, upscale. Of course, you have to type correctly, but then you get the corresponding results. That's double mouse click. The next shortcut is control-S. It opens the equivalent to save down here. That's the save button. You can save your workflow with it. It's probably very practical if you want to save in between. The third shortcut, and I found that a bit confusing, is not control-L for load. No, not control-L, but control-O. With this you can open the loading window for your workflows very conveniently and quickly. I find it a bit confusing. I would have expected control-S for save and control-L for load or something, but no, it's control-O, probably for open. That's the equivalent to the load button here. Maybe it wouldn't have made sense to call it open or something. But it's control-O. With control-D you can open the default graph. So a graph is here, the whole node structure. That's equivalent to the one down here, load default. It's actually always very practical if you want to put something together very quickly or want to have an entry point. Then you can just load it with control-D. With control-A you select all nodes. That's actually always the case. Control-A for select all is sometimes very practical, especially if you want to pull the nodes. But we'll get to that in a moment. So for the next shortcut I'm just going to build this default workflow. I'm doing a preview image here. And select another checkpoint. That's control-Enter. And with control-Enter you start the whole thing. So that's something like queue prompt here to press at the point. Is control-Enter. So maybe it can be very practical if you just typed something in here or something and changed a few values here. Just press control-Enter and the whole thing starts. There is another shortcut for this, namely control-Shift-Enter. And what control-Shift-Enter does at this point is, let's assume you have a series of generations already in the queue. Let's say 20 generations. And I press queue prompt and it starts. And you change something about these values here. For example, let's say we just want to reduce the steps. Then you can use control-Shift-Enter to put your just created rendering back here in the queue in the first place. That means it will be done next. It can be very practical if you have made changes and then just want to place them quickly in a long queue. So back to 20 and it's done anyway. Then we go to the next shortcut. And it's not there for that long yet. Although long is relatively October, November or so. But it was a long-awaited shortcut, namely control-Z and control-Y for undo and redo. So if I go here now and delete a few nodes, I can use control-Z to make the nodes again. This is sometimes a bit tricky, especially when it doesn't work so well. See if you're somehow stuck with the cursor in a text prompt. So you just saw I had to press control-Z twice until the nodes appeared again. I can delete them again. That's how it works. I clicked on the canvas again. If we delete them and the cursor hangs here, nothing happens. So select the canvas and then control-Z. And with control-Y you can make steps again. So that's how you actually know it from Windows and so on. With control-M, which I have already shown you a few times, you can mute a node. That's what M stands for, mute. This is equivalent to what you would say if you clicked on a node on the right and said mode and then on never. Then it becomes so transparent gray. And alternatively, mode always. You can also do that very comfortably with control-M. And that's very practical if you want to interrupt the workflow at a certain point. Let's say we have an upscaling here and you don't want to send every image you generate here as a base image to an upscaling. Then you can mute a node in the direction of upscaling and thus you only generate your base image. And if you like it, you can turn on the node again and then send this base image to upscaling, without having to send every missing image to upscaling every time. That's what control-M does. Alternatively, there is still control-B. And control-B stands for bypass. That means the node is still active, but its functionality is skipped. And we can see that, for example, if we now put a LoRA loader between our model. Then of course we have our model with a LoRA at this point, or our image generation. So I'll just let it run. Of course that loads a bit. Damn it, I'm still standing at 20. Wait a minute. I'll just press clear queue. So, now the LoRA is integrated into the image generation. If we now say control-M, which we had before, then the whole thing will make a mistake, because you see here, the case sampler says that the model is missing. Of course that doesn't work here, because this node is now muted, and that's why nothing comes back here anymore. The process flow now goes from here to there and simply dies here at the point. And for that, control-B, i.e. bypass, is suitable. I'll turn on the node again with control-M. Alternatively, you can of course also select bypass directly in the menu. But that's very practical. If you want to try the image generation without such an intermediate node, you press control-B, then it will be so slightly pink. And if you then press queue prompts again, then the whole thing works again. And what happens at this point is that the model that we send in here at the front is simply passed on to the output back here without this node providing any functionality. So in this case, in our case here, the LoRA is not used. And with control-B we can turn it on again, say queue prompt, then it will be loaded again. And from that point on we have our LoRA back in there. This is very practical to turn off certain things again, to turn on without having to rebuild the workflow. This also works, for example, with control nets or something like that, if you have a control net apply in between. Control net apply, where you apply a control net and then theoretically go over here from the conditioning. That will crash now, I won't load or start that, because all other things are missing. But let's say you have a canny depth map, whatever, control net integrated, then you can turn that off with control-B again and then the whole thing works again. So we can delete that once, we don't need it now. We go now directly again, is it still connected correctly? Yes, we can actually delete it and then we go to the next point. With control-click or shift-click you can add nodes to your selection. That means you select a node, hold control-pressed and then select the other nodes. This is very practical for shifting or grouping. We'll get to that a little later. The whole thing also works with shift, so either control-left-click or shift-left-click. A very important thing is coming now, you can copy nodes. And if you copy such a structure here, control-C, that's for copying, click somewhere and control-V says, then you make a copy of this structure here. But you see, nothing is connected here anymore. That means you have a clean copy that actually only knocks the cat on the door for a moment. Yes, I had to cut that, sorry. That was very important for the cat right now. So with control-C, control-V, only the connections within the copied structure remain. What you can do alternatively, we delete these three nodes, and you select something and you press control-C to copy and then control-shift-V. This has the advantage that, as you can see, all connections here also remain to the non-selected nodes from before. That means with such a setup, you can actually easily make different variants or to try out different ways. That's really super practical, especially if you're back here in the workflow later and you've loaded your VAE up here in the workflow, then you can also go and copy such a structure here, go back here, control-shift-V, and you can see that the VAE is still connected from the front here. If you then need another case sampler here, then you can easily connect it from here and don't have to go all the way back to the beginning of your workflow to connect the VAE here. Just as an example, of course, you can also connect the VAE with reroutes. Everything works, but control-shift-V is actually a very, very important command, which is also super practical, especially if you want to use the conditionings from before later in the upscaling area, then you don't need to connect them by hand, but you can simply do control-shift-V and the whole thing is already connected. The next shortcut is an ALT shortcut, namely ALT-C. With this you can collapse or collapse nodes and make them small. I can show you that here. ALT-C. Then it gets small and if you press ALT-C again, it gets big again. In the workflow I find it impractical to use it because you actually want to see what it does and what it is and so on. If we, for example, turn the case sampler small and so on, it's not super practical for the overview right now. One application would be, for example, if we go there and say we want to convert the text here into an input, then hang a primitive node on it with the text, because we want to use this story again at some point later, then we also have the text here, then we have a pretty useless area here, of course we can drag it down, that works, but in such an application case I think it is quite practical if you can collapse the notes with all c we pack the input back into the widget so we clean up a bit here and yes that works quite practically , by the way, if you select several notes then you can collapse them back and forth alternatively you can also double-click on it up here on this gray point here this one there then it works too but it's just practical if you have now selected several notes as c and the whole thing shrinks and can provide better overview here and there so the song we all know we take a note press remove or the song and it is gone with control z we get it back sometimes it lags a bit as I said the whole thing works by the way also with backspace is away it is back there with control z so backspace and delete do the same thing at the point if we now press control delete then the following happens then we are asked if we want to clear the workflow if I press ok now our workflow is gone again and that is equivalent to the clear button down here that makes control delete and control z and there is our workflow back there with shift and drag and drag means here with the mouse you can move several nodes at the same time so if I select several nodes here and hold shift key then we can move all the nodes that we have selected that is also quite practical to clean up the workflow a bit that does not work with control as you can see that works with shift so hold shift key pressed and left mouse button then you can pull the selected nodes so finally a few simple shortcut there would be for example the key q with q you can open and close the queue as you can see here and that can also be very practical if you work around a bit here and click around and then just want to open the queue once with q to see what the whole thing looks like here i still have 11 batches in here that was also wrong because you can quickly press q and clear queue again if you have too many processes in the q that is very practical with h you can open the history in the history you have here for example that is a bit similar to control z also so you have the option to create old runs again they are numbered so you can also go back to the steps that you have already generated in the previous process that is quite practical that works here you can see we still have our examples from before so I'll clean up again that works with the key h so you can open and close history of course also with the buttons here the q and the history but many then use shortcuts with r you can refresh once that is equivalent to the key here so if you have loaded models in the running operation and they do not appear here you can press r then come for a refresh and the whole thing will appear with you again and last but not least if you hold the space key the empty key I'll do that now I hold the empty key pressed and now I only move the mouse without clicking so you can move the workflow or your view of the workflow that works with the empty key and so nothing more click but simply move the mouse as long as you press the empty key you can move the workflow with the mouse as you want that were the shortcuts and in line with the empty key I would like to give you another tip if you hold the middle mouse button you can also pull the workflow then you have to press the middle mouse button and then pull you can of course also if you press the left mouse button I just do that work flow also pull but if you have really large workflows and you scroll out that the whole thing looks like this and you then start so imagine we have everything here full with some notes and it's just a huge workflow and you start pulling here then it can happen that you will come into such an input field here at some point here you see I got one that is now a drop down you notice it but it can happen that you also get something like the steps look here I got the cfg just when pulling it should actually be on 8 but it is now on 8.3 and that happens if you use the left mouse button as you want to pull through your workflow then you sometimes get such a thing and you wonder for a moment why why doesn't it work now it works again and then, however, it has already adjusted a value here and that can be very annoying at some point because you somehow have 5 or even 10 sampler or something like that here all the strengths of any control nets and loras and whatever built in then it can be very, very annoying if you accidentally change a value of it at some point and that's why use the middle mouse button because the middle mouse button works with that too I just press it so you can also pull but if I now click into such a field with the middle mouse button you can see that then you don't change the values anymore that's why it's best to get used to using the middle mouse button to pull or space and then pull with the mouse but I always use the middle one that is somehow more practical for me you have to get used to it a bit but it is really really more pleasant in the next point of this video now it's about updating the computer I know in the manager there is an update all function but it has already happened that it doesn't always work correctly so in my experience it works for the custom notes that you have installed, but it doesn't necessarily always work I haven't discovered a scheme yet I don't know why but always also on the base version of computer and that's why I would advise you to update your computer always manually and then start once and then update the notes and how that works we can now look here in windows portable goes into your update folder and uses the update computer but then such a window appears here it is pulled over the last version and the whole thing is then at least in the base version up to date if you want to do the whole thing in linux now go there and go to your computer and say git pull once this downloads the latest version and then you can start computer again and then we are back here at the next point because it doesn't matter whether via linux or windows then we have the computer manager and then you can say update all that then pulls you down the latest versions of all custom notes from git but you can also say fetch updates after that you have to update them manually I haven't done it today I update my computer every day I do fetch updates then we are told there is an updated version and yes fetch updates is not an update please update from install custom nodes that's what I just meant you have to go now and reinstall the whole thing manually when I press close again update all does that automatically in the background and the manager will tell you at some point your notes are up to date please restart once that goes in the meantime then over the surface here presses restart and the server restarts the server is in the case of ComfyUI but with the variant here you have a little more possibility if you look at it in the install custom notes directory and you update the filter up here then you can see for what there are all updates and that with a little luck then brings you into the possibility if you for example see here the impact pack has an update i look at the github page and see if there are any update notes here so notice is not notes with d but notes with t there you can then guess a little bit what has happened in this pack now alternative to that it is then maybe a good point up here on the right there is a button here so there is a number and then it is behind commits these are so-called check-ins of the developer so the updates that they have played in sometimes it is very practical to look in here because then you can see what has happened in the comments it doesn't always have to be something about the functionality with such updates or it doesn't always have to be new functionalities added it can also be that, for example, you are a windows user and everything works for you but maybe a mac user has contacted the developer and the developer has simply played a fix for max and you will also get the update so then it is also in your manager that there is an update for the pack but it does not affect you 100 percent or actually if it was mac error and no windows error 0 percent but that is always a very good starting point to see what has happened if you are a python fan you can still go in here and see what has happened in the code but I think that doesn't interest so many now but here these check-in comments are always very good if we look at the ComfyUI essentials, for example, because I know Matteo always has a very good documentation also up here we see the commits there we click on it and here, for example, we can see very well that Matteo has removed the stable123 batch node from the ComfyUI essentials because it is now available in the ComfyUI itself so ComfyUI anon has built it in this way, of course, Matteo has built it out of its own package, but of course that has the consequence that if you have saved a workflow before that has used the ComfyUI essentials that this node no longer works and if you want to know what happened there or something goes here on the commits and with a little luck the developers then also leave speaking check-in comments here so that that would have been such an update, for example, there Matteo fixed some bug you get in your ComfyUI just said hey there is an update thinks maybe super new functionalities or whatever no it was a little bugfix yes that's a good starting point to see what has happened here and then you can either press update here I would recommend you to do that for everyone who is suggested here or just an update all then again if you haven't done that from the beginning then you get the latest version I just wanted to show that there is the possibility of maybe finding out a bit about what has changed about the notes I'm just doing update all because I can't stand it if there are no pull changes at update all you can also see that in the console the whole thing rattles then updated all notes takes a little bit so now it's done and then you can press restart here ok then we see here in the console again that the ComfyUI server is restarted and as soon as it is back there we can continue working again with windows it is a bit different then a new browser window opens automatically then you can close the old one although it says reconnect and use the new one your last workflow will be loaded anyway with me it is now a bit different because I have installed my ComfyUI into the wsl linux integration into windows yes but you can see it yourself with windows it just opens a new tab by the way it is also important that you have your ComfyUI parallel the custom notes always keep everything up to date at the same time because I try to explain that a little bit we have a ComfyUI here at the moment he did it right now ComfyUI in version 1 well that's wrong thank you for that now ComfyUI v1 so this is variant 1 of ComfyUI and here you have now installed custom notes that are also in the now custom note v1 so let's make three of them they are all built in such a way that they can work with the current ComfyUI version if you do not update them in parallel now so your ComfyUI then so your ComfyUI and the custom notes then it could be that they run apart then it can be that perhaps the first two custom notes use functionalities from the old ComfyUI version this custom note here but in the meantime functionalities of the v2 needs so and if you then custom note version 1 with ComfyUI version 2 want to run then the whole thing does not work anymore that means you have to at least update this note to v2 the other way around the whole thing can also happen if you have not updated your ComfyUI here now but only the custom notes then you have this variant here and then it can be that the custom note no longer works and also throws an error so you actually always have to if there was any update variable it can be that there was a new version from ComfyUI then you have here in the variant there is suddenly a new version for this custom note as well as for the ones below then you have to update everything so that it fits again if there was no update for this custom note here then it will probably also work with the ComfyUI version 2 but what i want to say is updated your ComfyUI updated in parallel with the manager all custom notes that have an update available otherwise there can always be discrepancies between these versions and functionalities that are required by the custom notes have been changed in the basic ComfyUI or have been expanded or new functionalities have even been added in the new versions that are now also the custom notes then use and if you then have something like this that this custom note wants to use functionalities that are in the new ComfyUI version and you have your ComfyUI but not up to date then it doesn't work anymore that will throw you all errors in short updated your ComfyUI starts the server loads all custom note updates that exist for your setup installs them and you are always clean it often happens that questions come in the direction that and that no longer works for me or I have your workflow loaded and the error appears then the standard question is actually always is your ComfyUI up to date and are your custom notes up to date and very, very often the case is that there were actually discrepancies good the next point is drag and drop most of you will probably know it but the questions keep popping up again and again or it is asked for workflows even though a picture was posted and a lot of people are surprised, especially beginners, that you can pull in pictures generated with ComfyUI to get the workflow and an example of this is the ComfyUI page itself if we go on it and on examples then we can now take model merging then you can go grab the picture and drag it into your ComfyUI and you get the workflow that is behind this picture that works with generated pictures too I'll try to find one here in the painting there are generated pictures we just take that you take this picture pulls it into your ComfyUI lets it fall somewhere and gets the workflow with which this picture was generated with this safe image node this happens automatically I just have to see if I have installed the other one I don't think there are also safe image nodes that offer the possibility that you can turn it off that the workflow is called embedded, however, that is apparently still a lot of beginners as I said, that's why I think it's a very useful information to say that again if you see a picture somewhere that was posted many generate their pictures generate the pictures take the picture upload it somewhere high for example with us in the discord and then comes the question can I see the workflow for it then the answer is always in the picture then you can upload it by the way if you are in the discord that works in the discord but not via the preview in the chat but you have to go through the open browser functionality so in the discord is then open in the browser or in the browser open in German and then it opens in your browser and then you can, as I have just shown, drag it over and drag it into your ComfyUI and then the whole thing works you can also have pictures that wait a minute I'll pull that a bit to the right that are on your plate or something like that you can put the input folder you can just drag it into this load image node and then they appear a problem with it is you can already see he has now made a comma on one comma they are unfortunately duplicated somehow in your input folder that's a bit stupid if you copy a picture let's take johnny debb and then here with control v by the way then you automatically get a load image node and you can use the picture that just by the way well now I have to clean up a bit here again I go back again and then we come to the next point the next point that are the folders in itself that I have already heard that some do not know where ComfyUI saves the pictures or where you can get these pictures here in the input folder so in the load image just briefly at the edge in the ComfyUI there are several folders we see i am here in the ComfyUI folder itself and there is the output folder and in the output folder the pictures that you save with save image they end up in there you can also enter a folder here at the front then pack it up and the folders so then it is sorted out in folders there are also nodes that have they are then a little better set up I still have to reinstall a few nodes here after my switch they are set up so you can also enter a folder name so that you can organize yourself a bit then there is the input folder clicked the input folder here are all the pictures in it that are suggested for the load image node the load image node does not pay attention to subfolder the subfolder up here I made a few for myself are also created such as clip space or pasted created by ComfyUI itself but here you will find the pictures that are used for the load image node and as just seen in the example we have created copies here I can actually mark them and delete them once we do not need the copies that happened through the drag and drop here in the load image node here is definitely the place where you can clean up this folder here or this selection option once or just give your pictures a name that speaks differently than me here to make the whole thing a bit clearer for you yes i know i always wanted to rename the pictures but you know how it is another folder that can save your ass if we have a preview image node here now i do load default and say here i would like to have a preview image node who just restarted the server and we say you prompt ah so the anything model doesn't matter there is a folder in the ComfyUI called temp and in this temp folder, ComfyUI also lists the pictures that you have only provided with a preview image node or those that come out as a preview in the intermediate steps that are still there so if you have now somehow only worked with preview images so there we have our picture if you have now only worked with preview images and you think damn it now i haven't saved all the pictures at all then you still have the opportunity to go into the temp folder once i have to press f5 to refresh and here we see the picture that i have just generated here they are still there as long as ComfyUI is still running that means if you notice right away that you have generated more pictures but you have not saved them you can pull them out again here that works the temp folder is deleted when you stop the server and I would like to show you a server stop that you do with the command control c in the console so control c then it says stop server and the server has been finished and now we no longer have a temp folder here it has now been deleted under windows if absolute need prevails and you want to keep the pictures then you can also finish your command line or the console like this but I wouldn't guess that because you can see that things are still happening in a reasonable server stop so ComfyUI notices ok the server is stopped i'm starting to clean up again and leave my work environment clean and we don't know if other things will be cleaned up in the background from ComfyUI when stopping a server that's why it's good if you clean it up with control c once but as a small note the temp folder if you only shoot down the console then ComfyUI does not have the option to clean up this temp folder or remove it and then you may still be able to save your pictures and that saved me here and there if I needed youtube thumbnails or something like that, then yes, but finish your ComfyUI by stopping the server cleanly so I start ComfyUI again then the same thing connects here again and then we come to the next point if you have an Automatic1111 installation or even several installations from ComfyUI that is possible with the portable version of the protable version you can install as many portable versions as you like with different note setups or something or to try around so maybe one as a test environment one as a productive environment there are also one or the other in the professional area on the way with ComfyUI that works then of course it is only extremely annoying if you have to use all the models several times and normally you know that there is the models directory in ComfyUI and there is the checkpoints folder and all possible loras folders you know that I have already shown that but there is a possibility that you can change all ComfyUI versions that you have that they use the same model descriptions and for this there is an extra model path down here there is an example file and in this file so you have to rename it first then you have to remove the point example here at the back so just f2 and then remove the example I don't do that now because I already have it if we then say we would like to select it in the editor so then the whole thing looks like this and here you have the opportunity to either enter where your automatic 11 11 version is then ComfyUI pulls the models and everything it needs from the installation but you can also look at the bottom here the whole thing is commented out that is the point example file here in front is such a route or as we said back then latten fence here is a route and you just have to comment that out by removing so in this file and then you can enter here where your base part path to the models is and here you can name it as the folders are in the models folder also enter other folders and these paths will then be added when starting from ComfyUI and there you can also add custom folders as you can see below here I also have the ip adapter models are added they are not standard with ComfyUI but ComfyUI is so smart and then recognizes that he has to look for an ip adapter folder and on the given path and it results from the base path up here and from the path that you enter down here and the whole thing looks like a file system then I have created a models folder here and here are all the models in it that I use for ComfyUI and that now has the advantage that I can use them once for my windows portable version, but I can also use them like here for my linux version for linux then of course you have to enter the corresponding path of your linux path for windows the corresponding windows path but as you can see, ComfyUI is able to find all the models and everything even though they are stored from the ComfyUI folder itself, so here in my I'll show you my running version models checkpoints there is nothing in it nevertheless the load checkpoint node can find the models and that is because I have stored them here in the extramodelpath.yaml file so where we are in the folders right now, maybe you have already seen that I have a 15 slash in it and I have a SDXL slash in it plus a few loose things if you now look in my loras folder lora loader model only then I have the same here too we see a 15 up here we see SDXL down here and here I also packed celebs in between because these are actually the loras that I only use to get mixtures out of faces not the celebs themselves but mixtures that I had already shown in the video that I somehow mix 3-4 loras to be able to get rid of the trained models a bit in terms of face technology, the practical thing is then but that you are up here in the filter list if I now enter 15, for example, I only get all the loras that are for stable diffusion 15 I can enter SDXL and get all the loras that are only for SDXL but I can now also enter celebs and then get all the loras that are available for this face processing and the whole thing is basically super easy because you just go into your models folder there in loras and I just put subfolder on here we see the 15 again the SDXL I have something to test here and if we go into the 15 you can see the celebrities folder for face generation that is actually super practical so you can easily sort your models by packing them in the subfolder and ComfyUI picks up this structure, packs it up for you and you can then search for it very comfortably alone SDXL and 1.5 models makes sense to separate but you can also make other separations such as cartoon style realistic photography whatever I think you understand the point behind that it just allows you to enter SDXL much faster if you want to do SDXL and then you have all the SDXL models of course you have to start sorting them when downloading it saves you a lot of trouble and especially for beginners I think it's quite good that you can't handle models like that or don't know exactly what is stable diffusion 15 or so that is quite practical on CivitAI, for example, it always says right in a box which version that is, so be careful when downloading that directly to pack the right subfolder and you have a lot less stress in the ComfyUI to select the right models so what some may not know yet we come to the next point and that is how we do it now we don't need it anymore you can run several ComfyUI instances and the whole thing works just like that take the url up here copy it once opens a second browser tab says control v and you have a second ComfyUI running and the two can even work separately from each other you can only start a workflow but if we have a set up now, for example, where we now say we want to have a load image here we put the whole thing in a vn code that is already wrong vn code for in paint I wanted to have the example we connect the pixels we connect the vii is not quite clean i know we connect the mask and send the whole thing in here that means we don't need this image anymore then you can go very comfortably and generate a picture with the first work flow the picture then copy from back here then go over here again insert here and then provide a mask here and run an in painting and then save the whole thing here if you want to have it and always switch that means you don't need a big one build a workflow that can do everything in this case everything would be a workflow for generating a picture and then an in painting in this workflow but you can do it cleanly and separately just have a hand that tells me I now have a workflow to generate and then a workflow to be able to edit the pictures with in painting that makes the whole thing a bit easier on the one hand on the other hand but also much clearer so i just quickly clean up the control d to load a default setup again and there is since i don't even know how for me it is still relatively new although it is already two months old or so there is a node group function in confi and that is pretty cool and I prefer to make clear and you can, for example, build a setup at the beginning, which, for example, consists of a load checkpoint made of two clip text encodes that are prompting, let's say we want to add a VAE loader here to be able to exchange the VAE then you can now go and mark everything somewhere here in the empty area click and convert to group node then you can enter a name I'll call it loaders and we only have one node that combines everything and you can then also work that works now, for example, we can take a case sampler here we can connect the model we have our positive and our negative here we still need a latent image that means we can now go back here and say moment convert to nodes up here then you have to click on the node again that means we still take an empty latent image so select all the nodes again say right click convert to group nodes we give it the same name then confi once say attention there is already something like that yes we want to write about it and now we have all the values here again united in our yes in our group node can put the latent image in here and now we can choose our things here as we had we can say here we want to use this VAE this here we have two VAE outputs here by the way now here the one up here is the one from the loader checkpoint node and this one is from the load VAE node what is then here also a VAE loader VAE so with the name you can get along quite well here we can enter our latent sizes we can now say here acute cat text watermark and back here as usual continue to tinker with a VAE decode then we take our custom VAE from the loader and back here a preview image and then the whole thing works and that is quite practical to put together easier nodes it doesn't always work with everything I'll show you in a moment for something like that, but there is a small mistake with it, I just wait a moment until the thing is loaded so now it is loaded now we get our cat picture now the VAE is still loaded and there it is done that works quite well in any case what doesn't work quite well is that I have to check again the last time it was like this if we now say for example we want to add the width and the height back here and then I take our 1.5 aspect ratio from the comfirol aspect ratio and we can connect the width and the height here and we say now we want a 3 to 4 portrait of a cat then we see that somehow it doesn't work that well he still took 512 x 512 although we have entered other values here yes it is not quite round yet I think it makes problems if we store the input fields if we say for example we want to store our positive and say a cute dog let's see if that works that worked that with these comfirol things not then you have to be a bit careful that is always a bit strange nevertheless it is a pretty cool function I'll show you another example I want to go back to that if we now, for example, want to build an upscaling here at the back, I'll look for upscale with model loader then we take an image upscale with model we want to scale this image up we want to scale this image down again image resize let's say 1024 the proportion correctly slowly and at the same time we would like to take this code with us and back here so the code comes in here then at the point we also need a code so this pixel here takes no that was wrong we need a wii decode note in front of you which then later takes these samples here and from here we go with the picture in there so and back here we still need a wii encode note for the second sampler which then takes this one because we have to go back to the second one if I take a case sampler back here now we would have to be able to connect the latent image to it then that would be our upscaling setup at the point and if we go here now and say we convert that into a group note and say that is the upscale then the whole thing is a bit ugly and we now have two wii inputs up here that we actually don't need because we only work with one wii internally in this node so we could now of course say we are now connecting the story twice here but that is pretty ugly and you can then do a little bit by simply saying we are now doing convert to nodes again we take a reroute and connect them with the wii that are in here so now we have an input here that will go into two nodes later and we say right-click convert to group node we call the whole thing upscale again overwrite our previous setup and now we only have one wii input and yes that is that is already better because we only need to connect the wii once we have to look again where the image comes from we have to go back again we have an image input somewhere because I don't see where it comes from we have our latent input here we have connected that here we have connected that we have connected that it is a bit strange that he then says at the point we still need an image input but I said yes, it's somehow not quite round yet, now it's gone too, I just had a wrong thing in there, in any case, we have now built an upscaling node here or built a group of nodes that is quite convenient to use, you can throw everything back into the sampler we can now take the latent from here from our upscale node we say 0.5 the new ones behind we build a new decode node take the wii from here in front of the image and now we have a relatively complex upscaling process in a node summarized and the whole thing works quite well but always pay a little attention to that you that everything looks so correct and behaves correctly if you store inputs, for example, so it's not quite round yet and in terms of templates save and so I don't think it's quite mature yet but nevertheless I think it's worth mentioning so next I often saw the question or often more often that there can be problems if you now take such a setup here, for example, I just want another checkpoint with a setup where you want to generate several images within a batch and we know that you can do that down here with the batch size so now set the whole thing to 4 and let it run then we get four images in a run generated in automatic 11 11 it was as far as I know so that for each image that is generated in this batch down here, the seed is taken and increased by one that is of course better if you then want to make a picture from a batch again. ComfyUI behaves a bit differently at this point we have now got these four images generated from the batch so where it was with automatic 11 11 that we have a seed 0 here, for example, then it goes there seed 1 2 3 that is easy to reproduce ComfyUI works differently, as I said, the batch size is calculated into the processing of the image, but that also means we can't go there now and say we want to get the second image from such a batch made again by simply increasing the seed here or something that doesn't work because if we look at it now this is our second image from the batch in automatic 11 11 it would then be the case that the seed 1 would be appropriate but it is not there is but a trick that you can use here I'll put it back together and we have to make sure that we bring the batch size into the calculation of the sampler and there is a let me just look for several there at the very top I didn't look at it there is a ComfyUI a latent from batch node if you put the batch generation directly behind the empty latent image note then you could say here the index the index that is if we look at all of us that is here 0 1 2 3 that is a zero-based index so now it is at 0 that would be the image if we set it to 1 it would be the image at 2 the image and so on with the length you can say how many of these nodes should take from the bed so with 1 then we only get index 0 length 1 only stays with this image index 0 length 2 would be here 0 and 1 and so on index 3 would be 0 1 2 and that gives us the opportunity to say I would like to generate a batch of pictures first, that is, we take the full length that we also have in the batch size with index 0 and then we get the same results as before if we only want the second picture now this one then we can now say we want to have batch index 1 and length 1 and then we get this picture again generated that is then a very good or a very good way to pass workflows on where we don't always have to calculate the complete batch size but that you really say okay i have now had the four pictures from the batch from a batch generation but i just want to continue working with this one then you can set it here and then the sample only has to calculate one more picture we can also take the next one here that is with the let's say with the bit green in the background if we turn that back to 0 and length 4 then it generates four pictures for us again and that was that with the bit green background that is actually a very good way to control and say I have now generated a picture here which is very good but I have worked with batches before before passing on the workflows, the latest from batch node so that your target image is generated and then you can pass on the workflow and now the other users are not forced to calculate all the batches over and over again, I mean with four is still quite good but imagine you have ten pictures of a large batch or even more that depends a bit on the hardware that can then be quite annoying and then when people with smaller hardware want to generate the pictures then it may not work and they don't know which picture you have passed on from a batch that is larger so the penultimate point is already quite long I have to see after cutting how long it will be the next point is I am often asked how I get these these render modes here so I mention that again there is in the computer settings that is here this little this cogwheel this gear here there is a link render mode you can easily change it and then the renderings change, especially beginners don't know that here you can change it very well very comfortably so you have to pull once that is now the classic view I have always had the linear view at the moment so that I heard it again very briefly on the edge and the last point in this video is just a request to you how you can deal with developers I get that more often now via the discord and so and matteo is a developer here and we always work very well and closely together and it is so that many post a screenshot of the error message that always appears in red here and that is always very difficult for developers to understand, so if you have an error, go to your console, mark the error text that is in there and copy it and pass it on to the developers, whether it is now in any discord or in github or wherever you don't work with screenshots but copy paste the text that is in the console you can also get the text out of the screenshot or out of the error message here in the computer but if possible so please don't go and take a screenshot of it and give it to the developers but give it in text form that makes it much easier for us to develop custom notes for people who work hard to find the errors you may already have a small resolution jpeg compression when you are alone in the screenshot it will be difficult to read you have to look first maybe it helps one or the other developer then to take the text to format this message a bit with line breaks so that it is easier to read and so on, so don't make a screenshot in a short time but copy the text variants yes and with that we have now reached the end of this video i didn't think it would take so long but yes there were now a few collected works from the last two three months that I always noted next to me and that I wanted to show you and yes it was worth it now to make a video for it, let's see if I maybe in between I always bring videos like this depending on what comes up with questions where I notice the questions appear more often that you can maybe edit them a bit here and show and discuss and I hope it helps you in any case I would say we'll see you in the next video again which will be a bit more technical but until then yes I hope it was helpful for you, we'll see each other and bye
Info
Channel: A Latent Place
Views: 2,070
Rating: undefined out of 5
Keywords: ComfyUI, Stable Diffusion, AI, Artificial Intelligence, KI, Künstliche Intelligenz, Image Generation, Bildgenerierung, LoRA, Textual Inversion, Control Net, Upscaling, Custom Nodes, Tutorial, How to, Prompting, Tipps, Hints, Advices
Id: lFDb9MZ72sM
Channel Id: undefined
Length: 68min 44sec (4124 seconds)
Published: Mon Jan 15 2024
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.