Hello and welcome to this video, in which I would like to exchange some life time against knowledge. I took a look at the SDXL Turbo Models, which are now sprouting from the Comunity. I wanted to know if we also have the massive advantage of one-step generation. Unfortunately, this is not the case. Nevertheless, these things are still super fast and also deliver better pictures than the base SDXL Turbo Model. Before I show you how you can find these models, I would like to start with a small announcement. In the last few weeks I chatted a lot with Matteo and also called him. If you don't know Matteo, he owns the YouTube channel Latent Vision and he is the developer of the IP Adapter Plus Custom Node Packs. I have already presented this on this channel many times. We decided, because Matteo was a little worried, to open a Discord channel for his activities alone, we started a collaboration. The Latent Place Discord has become the L2 Discord and serves as the starting point for our two YouTube channels, Latent Place and Latent Vision and also as a support area for the IP Adapter Plus Custom Nodes. I personally hope that we can create a good community that can exchange knowledge and help others, be it beginners or advanced. That everyone has the opportunity to get into this topic of AI. So jump on it, look past it and we'll see where the journey goes. Well, now I'll show you how to get the best of the SDXL Turbo Models. For this we have to go to civetai.com, which should be known to you. If not, you know it now. 1a starting point for all kinds of checkpoints, loras, embeddings, whatever we need to create the AI images. There are also workflows for ComfyUI, open poses, actually everything. When you are on the page, go to Models here at the top left and then you have a filter option on the right side. Click on it, select Checkpoint here and down here you can select SDXL Turbo. And then you see all available models or checkpoints that are based on SDXL Turbo. And if we sort this again to most downloaded, I did that. Then we see a pretty good selection of all the checkpoints that we now have available. Overall, however, it is also relatively limited, because that is all there is at the moment. Nevertheless, there are good ones. I took a look at the DreamShaper XL, which has a Turbo variant on the page. Let's take a look inside. I took a look at the Copax Timeless XL Turbo and the JibMix Turbo XL. I downloaded them and tried them out. And these are the three. They make really good pictures, at least better pictures than SDXL Turbo in the base version. However, this also comes with a bit of performance loss, although you really have to say that. That's already complaining at a high level. I'll show you that in a moment. Down here in the descriptions of the models we always see what the creator suggests for settings. And here we can already see four to seven sampling steps. Here we have one to four steps. And here we have number of steps five. You should take that to heart at this point. Use that as an initial value and then look at your desired result from it. That's why I know about it. So if you start using the SDXL Community Turbo Models here, then please look in the description and stick to it. It is also interesting that the DreamShaper XL is apparently a trained checkpoint, although I don't quite know now. They change here, that stays the same. He is trained, the other two are merges. So checkpoints were merged with each other, mixed. And then I'll show you later in the video how you can do it yourself. But we'll jump in now. I also had a viewer comment who asked me to put the connectors on linear, because that might increase the overview in the videos here. I heard you or you and will try it out. Let's see how that will be in the future. Maybe I'll switch between the connector styles. Depending on your taste or so. I don't know yet. Let's see. But let's start now. I first load checkpoint and as a reference I take the SDXL Turbo Base Model. This is what I presented in a separate video. I make the note a little bigger because otherwise you can't really read the checkpoints. And from there we go into a sampler custom. So I'm going to build the SDXL Turbo Setup here again, which was presented on the ComfyUI page and by StabilityAI. So we now connect the model with the model. And what I just said is the connector here, I had it on linear before. That means it was about corners. Now it's straight or I think it was called square. You can change that here in the settings. Link render mode. Yes, now it's linear. It was straight before. And there were problems when I do it that way. Or at least that's how it was noticed. And unfortunately I have to admit that if I build something like this now and we say we pull the sampler over. And I somehow built something like that. Then you have overlaps here and you can't really see what's going on with it. But if I switch that to linear now and update it once, you can already see the connections much better. So I heard in the comments and we'll just do it now. Let's see. So we definitely need a negative prompt, which I didn't know at the time of my Turbo SDXL video. Negative is completely ignored by Turbo SDXL. That's why we'll just make it small. The sampler still needs to be sampled. But no matter what we enter there, it will just be ignored. That's why we just make them small. And just so that he doesn't rot here, we put them in. As a sampler, we take a case sampler select. And here we can switch to, I would say, SDGPU. Let's take it. It's a little more random. And as Sigma, the SDTurboSchedulerNode was entered here. Since Stability AI, we can enter the steps. It still needs our model. We can pull that up here again. Then we need a Latent Image. For the Latent Image I always like to take the aspect ratio from the ComfyRoll suite. And for Turbo SDXL we also take the 1.5 aspect ratio. Because Turbo SDXL is trained to 512 and not like SDXL to 1024. So let's take a portrait picture so that it looks a little nice. And back here we take a VAE Deport and a Preview Image. So we still need the VAE from the checkpoint up here. And that should already work our setup. I'll do queue prompt once so that the model is loaded. On the subject of models. I had all three models that I just mentioned loaded on top of the SDXL Turbo model for testing. That works too. However, it becomes very tough and very uncomfortable. And I noticed while recording that this is already the second attempt that I record this video. That I can only load one more model. So we can't put all four against each other. Because I just need the RAM for that. I have 32 GB of RAM in the computer. And the models, when they are loaded here in the load checkpoint, land in the RAM. And when they are used by the sampler, they are pushed into the VRAM by the RAM. But if the RAM is full now, it is stored on the plate. And if the RAM is full, we know that. Then other programs such as OBS react, with which I no longer record my videos so nicely. And yes, in addition to the fact that after a full RAM you still have to be loaded from the plate, the whole thing becomes extremely tough here. So we can load one more. We can switch that too. Unfortunately, I can't put all four models next to each other in this video. So we have our first picture. And I can now say ... ... a cute cat. So we have to set the auto queue in the extra options. Oh, we still have to adjust the CFG here. By the way, that's something I always like to forget. I like 1.1 best. And we have to set the seat to fix. So now I do Queue and because I have set the auto queue here, something is always added here. That means we can now, as we did in the last video, just type ... ... a cute cat in kitchen drinking milk. Now we have our cat drinking the milk in the kitchen. And now we can easily go through the seeds here. That's cute, isn't it? Let's leave it. And then we continue. Let's take another load checkpoint node. I copy that. And now I take one of the downloaded Turbo Models. Namely the DreamShaper XL Turbo DMPPSDE Model. And now we need a normal sampler at this point. Because we ... Wait, I'll take that out. He doesn't have to do that right now, the auto queue. Because of course we have to take into account the creator information. I connect positive and negative from up here so that we only have to operate a prompt. We take the same latent as we had in the SDXL Turbo. We store the seed from both here. Noise Seed to Input. And here the Seed to Input. Make a double click on it here. Then we have our Noise Seed here. We can connect it with both. And now the whole thing is synchronized. Down here for the sampler. Now we have to see what the creator suggests to us. The CFG Scale 2 says 3-4 for style stuff and 4-7 for sampling steps. We are supposed to use a DPM++ SDE Karras setup. What do we do? We take a DPM PP SDE GPU and a Karras. So we are also synchronized to the one up here. I had also set SDE GPU. The noise stays at 1 and steps between 4-7. I would say we take 5. And the CFG says 2. I have to look again. 2 he says fix. So not from bis but 2. And now we look at the whole thing. I make the node here bigger so that we have more cuts. We copy a VAE Decode. We take the VAE from up here. And that was stupid. I wanted to do it that way. I thought I had already copied. We take the one in there and the one in there again. And then we have that. To stay clean here, I would also say we make a storage of our positive prompts. And that is. Take that out once. Say here we want to input text. Make a double click on it. And take that as our prompt. Here I add our old prompt again. So we can make the node smaller here. Pack them into the negative. And copy them once. We also take the text from here. But the clip from this node here. So from this checkpoint. And now we have to connect it here again. So it's just that we can also represent the clip cleanly. Now let's run the whole thing once. Now of course the model has to be loaded. And that is at the time when we have both models in the frame. You can see that well in Task Manager. The frame is filling up very quickly. And that's what I noticed. Especially with the SDXL Turbo Models. You are very quickly at the limit. So it starts. And there we have our picture. And now let's take a look. I have Auto Queue on. And it still works. We are now getting both pictures updated. But you can already see that it is getting pretty tough. Now we see here that the picture is not yet quite optimal. And now we can start to screw around a bit here. He said somehow more steps for styling. Let's do it. Yes, six steps looks pretty good. We can still adjust the CFG a bit. Let's see. Until it somehow fits this motif that we want. I take 1.8. I actually like that very much. As a base picture. And now we can compare a bit. And we can already see that the trained models actually give better pictures back. Or the models created by the community give better pictures back. However, at the expense of a bit of performance. Because of course we no longer work with one step as stated in the Turboscheduler node. But with six steps. And now we can change the Seed here. And see what differences we get. So it does help to look at these trained or merged models. To get other results. I can now also say again here. Blonde woman sitting in a cafe in France drinking coffee and having fun. Not gun, oh god, fun. Oh, bad con man. Oh. Oh, he's still on it. Wait a minute until he's done. So. Yes, and here we see again. The differences between the models are not yet perfect. I think that the motif has changed a bit here. We should start tweaking here again. But he is actually a very good example of how it shouldn't be. And that we are going in the better direction. I'll turn this one off again. Then we can adapt a little bit to CFG here. Now it's faster. After ComfyUI know he doesn't have to worry about the upper path anymore. We'll see what gets better here now. Yes, less CFG. It's getting a little better if we put it on two. Could then try Seven Steps again. Or five. Six is already a good average. Let's take 1.2 or so. Yes. Maybe the seed also misses. We can adapt it again. In any case, you see what I mean. There are already more beautiful pictures coming out. And we can now also just here. Also relatively quickly. Create pictures. Acute Cat. Sitting, not Whitting. Sitting in a field of flowers. Summertime. Trees in the background. So that's already very, very fixed. Then you don't really notice whether it's one step or six. But the pictures are getting more beautiful. If you run the same thing again with our other sampler. The error just came because nothing was activated here. We can ignore that and just click away. This is the difference between the prompt with the same seed and the settings. And I have to say, you can see quite clearly that it is better here. We take it away again. Ignore the error message. Take it back in and say we want to try another model. For example, the JibMix Turbo XL. Be careful, a bit like the Whiskey Mixer. That I'm not saying anything wrong there. Now it's loaded. As I said, I think Turbo SDXL. In terms of performance. So from generating the pictures top. From loading and handling in between. At least the others are still a bit further ahead. And here we have the problem again. Our settings no longer fit. We took the JibMix. What is recommended here? From the creator 5 Steps CFG1. DPM++ 3MSDE. DPM++ 3MSDE. 5 Steps CFG1. It doesn't work at all. Let's take the GPU. Not either. Yes, there are already very strange results. No, that was the wrong one. Sorry, that was the upscaler settings. You have to pay attention to that too. We are supposed to use Euler A at this point. We can try that again. An Ancestral Sampler. See if it makes it better. Yes, thank you very much. And that's why it's important that you look at the pages here. And find out what the creator of these models advises you. To be able to work with it. But now we can take our other motif again. That looks pretty good actually. Let's take a look at CFG1. What changes? Yes, that's a little better. 1.2 And now we see Blonde Woman. Riding a bicycle. In a park. Summertime. Smiling. Yes. For comparison, let's change the prompts again. The sampler, sorry. And here we see again the advantage of the trained model. In contrast to the SDXL Turbo Base. I think that's pretty clear. While I was looking through these pages, I asked myself. How can it actually come to that? That we now need more steps again. And the information is all so different. And that led me to the question. Can we actually create turbo variants of SDXL models ourselves? And yes, that works. I'll just turn off the auto queue here. And we'll start again from scratch. So that my frame is a little emptier again. Yes, that gives me a little more secure feeling when recording the video. I hope nothing went wrong anyway. Well, let's do a little setup now. Let's start with a checkpoint loader again. And we load the SDXL Turbo again at the point. So. Then we take a second checkpoint loader. And let's leave it as it is. And now take a model merge simple node. And connect the two models together. And here we can already set a value of 0.5. This means that half of the two models are mixed with each other. And that's actually the technique that describes when it says checkpoint merge. This means that a new one was generated from several models. And then linked with different strengths and weaknesses of one or the other model. From here we go into a case sampler. Because we have to go back to the adaptation options anyway. We take a case sampler directly. We take a positive prompt as we had. As we had a negative prompt. But we ignore it. And I just take the clip from the Turbo model. I don't know which clip variants were used where. But our goal is to create a Turbo model here. And that's why we take the latent image. As I said, I'll take the Comfrey Roll aspect ratios again. Set them to 3 to 4 portrait. Take the latent image. And back here, as I said, to take a look at the whole thing. A VAE Decode. Then I just take the VAE from down here. We can make the preview image a little bigger. It won't be quite 3 to 4. Maybe something like that. Now let's go back and say Acute Cat. And let's run the whole thing. Let's see what happens. Okay, the whole thing is working. Wonderful. Let's go to the front now. Because we have now loaded our SDXL Turbo model up here. And down here we load another SDXL model. And I would just say we take Juggernaut. And then go in slowly. Our goal at this point is to generate a picture with as few steps as possible. And that's why we turn it all down again. We take steps 1. CFG 1. And take another sampler directly. Let's stick to the specifications of the other creators. And see what happens now. Of course I still have to load the Juggernaut here. The SDXL models are somehow less resource intensive than the Turbo model. That's my feeling. I looked at the task manager performance a bit. That was already bad with 4 models. At least 4 Turbo SDXL models. We got a picture. That should be fast now. And now we can take a look. If we do this setting with the Turbo model. That means we turn it to 1. Because then it takes 100% of input 1. That's our Turbo model. Then we get the picture out. There you can already see that the Turbo model generates with only one step. But now we can start to slowly mix our other model in. And there it goes a little bit. That we have to tweak a bit. But we are getting closer and closer to our result. And if we go in with 0.5. We have a 50% mix of the model again. And depending on how strong our model should be here. From there we start to tweak these settings here. I turn on Auto Queue. Say Queue Prompt. And now we can tweak that comfortably here. And we'll see what happens. We can already see that the more steps we put on it. The better the picture will be. We can take a look again here. If anything happens here. We even need one more step. I doubt it. Maybe we'll just go in with 6 steps. Let's see what happens when we turn the CFG to 1.5. Whether it gets better. No, lower CFG is good. Maybe 1.2 or so. Now we can start again. Maybe a little less. To mix in from the Juggernaut. For this we have to turn up the value here a bit. Let's see until all the changes have gone through. Well, before that it was better. Maybe less Juggernaut. And so we can get a little closer. Or less Turbo SDXL. Sorry. So we can get closer. And see that we get our desired result. In any case, that's what I think these recommendations are for. Depending on how they were merged. You take the Turbo SDXL model. You take a SDXL model that you want to have. Mix them together. And then start to tweak a bit here. We can look through the samplers again. Well, it's worse. We'll take the 2MSDE. Also works. The SDXL GPU is also okay. Let's leave that one. What you can still try at the point is, if we take a load VAE under loaders. And put that in the decoding. There is a SDXL VAE. Let's load that in. But it may also be that it has already been baked into the Juggernaut. I don't know from the head. But we can do it that way. It definitely doesn't hurt at this point. And if you want to have the whole thing the way it is now, then we'll go there. I'll turn off auto-tune. I didn't catch it now. No, focus was somewhere else. We turn them off once. And then go from back here. And then we go from here. And then we go from here. We turn them off once. And then go from back here. Add node. Advanced. Model merging. Checkpoint save. And here we can pull in our checkpoint. Here we can pull in our clip. And if we're not sure about the VAE, we'll just take it with us. And slap them in here too. And call the whole thing now. Turbo. Turbo does. Because we're doing a tutorial right now. Let's call it Turbo does. So let's run the whole thing. Then this checkpoint save node rumbles in front of it. Now it's done. And I'll show you where he finds them. Namely. If you. Look in your confi UI folder. ComfyUI Output. Is an folder called Checkpoints. And if you look in there. Then we have the Turbo does here now. 0000. 01. Safe Tensors. We cut them out and go into the ComfyUI folder. Go to Models. Go to Checkpoints. And. They slap us somewhere. I've got a little bit of it here. But now I'm just going to slap them into the Models. Checkpoints folder. And that was it. What we can do at this point is. We don't need the checkpoint. Safe node anymore. We don't need the Model Merge node anymore. We don't need the VAE node anymore. We can activate the whole thing here again. Control M. We delete this. Checkpoint loader node out. And we'll do it here. Refresh. And now we should be able to. Our Turbo does. 0000. 01. To load Safe Tensors here. We have to connect the VAE to the VAE Decode again. And since we haven't changed anything now. Let's run the whole thing. Now, of course, he has to. Our latest model. Load into the memory. That's what he's done now. And we have our. Sweet little cat picture again. And now we can go over here. And say. Auto Queue. We say Queue Prompt. Of course, nothing is happening. Because nothing has been changed in the values. But we can say now. Soldier. Holding. Gun. Or. Woman. Sitting in. Café. In France. Smiling. Of course, you should write it right. But also the model is so fast. That it is right. Writing error very quickly. Yes, forgive. And we still have the option. To adapt our CFGs. And so on. Or the steps. I know I just twisted it. Then you have to. Certainly invest a little more time. As we have just done it. Here in this video. But I think you have the basic. Thoughts. So you take the SDXL Turbo Model. Merge it with the Model Merge Node. With another together. Divide. You can play around here a bit. Whether that fits or not. And then you gradually. Just. Merge your own models together. Which are very fast. Are able to generate pictures. Dog. Barking Dog. Yes. And here it is actually. On the strengths. Which the model. Which you mixed in anyway. Had brought. But the dog is with our settings. But again. Pretty good. Running on the street. Yes, I think so. As base pictures. Work very well. Now a final tip. If you now. For example, the performance. So want to use the speed. The performance. To make several pictures at the same time. I just want to give you. And you build. An upscaling process. So now here at the point. We have a batch of four. I just have that down here. Increased in the batch size. We have a batch of four. And that means we get now. Swoosh. As I said, at the speed you can. Actually not here somehow. You can't complain about it. Or something else. Come on again a little bit. And you build an upscaler in the back here. NN Latent Upscale. For example. And we say now we want the whole thing. So big. And. From here it goes again. In a case sampler. Let's take it like this. We take that. Model is actually. Leave it to you. We could add a checkpoint loader here again. And say. We take the right juggernaut here. Then at the point. So. Pull it all in here. However, we take the conditions. From back here. So from up here. After this error. Again and again. It's so convenient to just drop. We take Latent Image. From the front over. And say here the noise. 0.5. But then here too. Adjust the steps again. That's just normal upscaling. As we know it. But we can't do that. All the time. Let it run. Our prompts enter. And. Here is a good opportunity. To say something like. At the moment we would. All four upscale. But there is. A note that is called. Now I just had to look again. Excuse me. It's called Select Latent from Batch. It comes from the Quality of Life Suite. It is also linked below. And with that we can now. For example, say. We send our Latent. From our Sampler. After that in our upscaling process. I have now taken a NN Latent upscale. And then we can. Look. We do the whole. Path here now. Just silent. So control M. Mute. And. Let's have another picture here. Generate. Or four pictures at the same time. And if we say that down here on the right. I like that. Then we go there. Make us the things here. Say here of course. Again, so that we can see something. A VAE Decode and a Preview Image. We also see bigger. So. We want our picture now. Upscale down here. Now we can say up here. And from the batch. This is picture 0 1 2 3. So let's put this one here on 3. We chase the whole thing again. Into the queue. Of course we have now. At the point here a normal. Upscaling no longer the turbo model. But a normal upscaling. That's why it takes 20 steps again. But we see that we are now here. Or that was before. Because I had auto queue on here. I'll turn it off. We see that we are now. Our picture here once. Have upscaled. To a decent size. So we can speed up here. Use. To let us generate pictures. And when we are satisfied with the pictures. We can chase them into a normal. Upscaling process. We can also expand the whole thing a bit. Because there is. In addition to this select latent from the batch node. There is another one. And that means. No, I think it's better if I look in here. Because it is under latent. The latent selector. So we can. Choose several. That means we hang them around here now. So. And we say. We want. From this set of pictures. Picture 1. And 4 have. That means we have to enter here. 0.3. Let's run the whole thing. Now we have here. We only see the picture now. But we can close it. And we see. We have. Once this here. And once that. Error on my part. We try it again. We take now. 1.4. And now it's right again. We have now got the picture once. And once that. You have to be careful. This node here is not zero. That means he really takes. 1.2.3.4. And he calculates here. 0.1.2.3. I was not yet. So aware. I learned something again. Very good. And if we want to have new pictures up here. I'll take them away. Then we only need this one. To mute. And we can run a new one here. Seed. I don't have an auto cue. It doesn't matter. If we have generated pictures here. Or we say. Now. Cat running on the street. We get super fast. Generated four pictures. And here we want. Now times. Two and three. Have enlarged. That means we turn on this emergency again. Enter 2.3 here. Let the whole thing run again. And even if it's slow right now. Here. You have to be clear. Two pictures are currently being scaled up at the same time. That's why. It's already possible. Especially if you consider. We're doing that now as a preview. As base image generation. And then we can select. And put them in the sampler. So we had two and three now. We now have this picture here. Number two and number three. Get again in larger variants. We can also switch down here. And yes. They've really grown up. So it's not a nice one. Scaling process now back here. It's relatively simple. But in principle I just wanted to show you. That you have the speed of the turbo models. Then you can use it again. To create several base pictures. And then the desired results. Either individually. With the selector node. I just showed you. Or with the computer image selector. Node. In your upscaling process. To send in and. Do you find this need? I'll show you that too. Quickly. If you are in the computer manager. And the selector is looking. Then this is the computer image selector. Node collection. It doesn't come with very many. But with useful. Nodes. And there is once image selector. Image duplicator. And latent selector and latent duplicator. So you can make several out of one. Or from a batch of images. Or latents. Pull out what you want. And so that you are not completely lost. Here is also always included. Where you can find these notes. Namely once in the image folder. Of ComfyUI. And once in the latent folder. Of ComfyUI. Well yes. I found it very interesting. That there are now these community models. That we can use them. Also interesting to know that they. Not quite the strengths. Includes the SDXL. Turbo brings with it. But can generate much better pictures. With a little. Performance loss. Maybe from one step. Up to five or so. Nevertheless. It is still very, very good. And very performant. I showed you how you can do it yourself. How you can save it. And can use it. And can also use it. To always generate several pictures at the same time. And then quite comfortably. In an upscaling process. To send your choice in at the end. Well. That was unfortunately a long video again. But I hope. Yes, I was able to convey something. I wish you a lot of fun. When building yourself. And experiment. And try the. Community models that exist. Or when creating your own. Models with the Turbo variant. There is a lot again. Possible. I hope to see you again in the next video. Take care until then. Bye.
