Hello and welcome to this video, in which I want to exchange some life-time knowledge again. The IP Adapter shares. This is again a video from the series of IP Adapter Updates, now in version 2. And today we take a look at the IP Adapter shares. This is also a very nice thing and it is best explained when we simply load a default. I say here in the back as a file name to save, I take the IP Adapter shares. We put our sampler on the correct values. I take fixed, I take a CFG of 6, DPM PP2M, Karas. So here in front I choose an epic realism and we leave the prompt empty for the time being. Well, down here I still take from the ComfyRolls, where are they? Where are you there? Once the aspect ratio and we can then adjust it depending on which image we take as a reference. So the IP Adapter shares. Here it is. We need a unified loader first. Of course, we can also connect everything by hand if necessary. But I'll stay with the unified loader. I take the plus model. We connect the model here, take our model from down here. We already know that from the previous videos. You are also welcome to take a look if you haven't done it yet. Connect the model to our sampler. I go directly with the weight a little bit down. And then of course we need a reference image. So and here I take the boat on a lake from Pexels. And we can see here in the IP Adapter shares already. It still has two more outputs. They serve as a reference, namely the tiles and the masks. And that's what I meant. I'll do a preview image and a mask preview. If I let that run now. Then we can see very well what the IP Adapter shares. So now we can also switch to landscape for a moment down here. I take 4 to 3 landscape is probably a little better. Yes, so the IP Adapter shares. If we take the Prep Image for Clip Vision node. We can choose whether we want the left area or the right area of this picture. We can look at that again by doing a preview image. So now we have picked the left area here. Of course, we can also pick the right area. And that takes over the IP Adapter shares for us. And we can see that up here very well. So I make a little space. We see that the image, which is in the landscape format, has already been divided into two images. And we also get the right masks for it. That means this mask for the left side and this mask for the right side. And now the IP Adapter shares that the descriptions are applied to the respective image areas. Which he also gets from the respective image areas. So left in the picture or right in the picture. And that is the strength of the IP Adapter share. We get very similar pictures again. And of course we can also adjust everything possible here with the other IP Adapters. Prompting and so on, we have other examples in a moment. The whole thing works if we say we want to load the Mona Lisa. And we now take a 4 to 3 portrait for StableDiffusion 1.5. 3 to 4, sorry. We'll let that load now. Up here we can already see that we get the descriptions or the image separated. And then also individual descriptions that are applied to the masked areas. We take another Seed. We'll take a look. Yes, a very similar picture. If I switch to the Mona Lisa Crop now. And let that run. Now we go down here again to 1 to 1. So. Then we also see an advantage. This picture here, the original picture, I'll open it up. Is 2000 x 2000 pixels. And the IP Adapter shares that with us. In four pictures. We see it here. We can also go in like this again and we see the four sections. A bit problematic is of course when you have such an section here. Where little description is available. And we can push the whole thing a bit. By hanging another IP Adapter behind it. So, I'll just take the model from up here. I also want to have my Plus Model here. The IP Adapter Plus Model. That's why I take the IP Adapter also forward. I'll do Prep Image for Clip Vision here. Because we need it at this point. Now we need it again. Because we have a 1 to 1 ratio in the picture, we need it again. So we don't need to share it anymore or something. I still do a Prep Vision. We can put a little sharpness on it. And we connect the model down here. And we take the strength down here. From our IP Adapter Advanced. And let the whole thing run again. What we now receive additionally is a description of the overall picture. While the individual sub-areas are described here. You can also see it very nicely. We had a 2000 x 2000 large picture. This was reduced to areas that are 512 x 512. And in these 512 x 512 areas we have small pictures. Which are only 256 x 256 in size. And we have seen that the whole composition has become somewhat more stable. Because we have added a second IPA here. Of course, we can also do everything here. What we can do with the others. For example, IP Adapter. I'm looking for the noise. Where are you? There is the IP Adapter noise. We can add a slight noise here. I take gaussian. Well, that's how they hang into the Image Negative. Let the whole thing run again. And get a slightly sharpened and, I like to say, more stable picture at this point. If you ask why I put a Prep Image in between. Of course, the IP Adapter also takes care of the rescaling. So to scale down from 2000 x 2000 to 224 x 224. But with the Prep Image for ClipVision we can still choose the interpolation here. Lanczos is the best. And of course we can put a sharpening on it. That doesn't happen in here out of the box. That's why you can still hang this note in between. So, let's start with the prompting. So, I'll do a few security terms in here. So that nothing goes wrong. And I would like to have a zombie woman standing in front of a graffiti wall. So, let's run the whole thing. And there's not much going on. Zombie seems to be triggering a bit now too. But it's not doing that much now. Of course, we can switch on or off all kinds of things here. We can also say Reverse In Out here. Let's see what happens. Yeah, doesn't change much. We can try it here too. It's being driven by the model again, as you can see. I just want to show that we can still handle the weight types here as we want. But what I want to get out of it is. We still have, even if we get masks generated here from the IP Adapter, we still have the opportunity to use our own extension masks here. And if I edit that in the Mask Editor now and say, I just want to have the area of the lady here. Our Mona Lisa. Then I smear a mask over it quickly. Say Save to Node. Take a mask blur in between. It is always advisable to take mask blur to make everything a little softer. And we'll put this mask blur in our attention mask. And let's see what happens then. I turned back the weight types. So that we have a little more play. So, and now we get more of our description of the Mona Lisa in the middle. And the graffiti wall in the background. And what is really cool to observe is that the masks of our IP Adapter share the created masks. Also arrange with the attention mask at the same time. You can see that very nicely here. That means we are now getting part descriptions from the respective areas. But we also get a general global description of the picture through our second IP Adapter Advanced. So, I'll just look around again. Maybe we can push the colors a little more. Yes, that looks pretty good. And now we can also play around here. What happens if we use the noise from our original image a little bit? Let's see what it makes of it, whether it gets better. No, it's a little worse. Generic noise seems to work better there. Let's go back to the picture and look further. The IPA shares is also great for getting a little stability into the upscaling. And for that we do the following. I'll take a resize node. And let's say we want to enlarge our picture to 1024 pixels. We take a second sampler. No, I'll do that differently. So that the connections stay. We want to have the prompting. We want to have a VAE encode here. For that we need the VAE up here. I'll just do a reroute here because I think we'll need it again. So push that in there. And say we want to have a denoise of 0.5. Now let's take the VAE decode again. That's why I put a reroute here. So we can now take a second IP adapter and say we want to put the original model in there. I still take our IP adapter from up here. But as a reference image I would like to have our generated image from down here. Then I want the strength to be one. And we send the whole thing into the model of our sampler. Parallel to that, before I copy, we should use another seed for upscaling. Let's take 6. Now I can copy that once. And the difference down here will be that we use the normal model for upscaling here. And to compare the results, let's take an image comparer from the RG3 node collection. Oops. So and so. Make the whole thing a little bigger. And let it run. So, it's through. And we can now see that this is the variant that we got through an IPA. And this is the variant that we got through the normal upscaling. And if we now take the reference image with us. That was this one. I'll make it a little smaller. Then we can actually see quite well that the upscaling here, of course we have changes in the image. But the upscaling has oriented much more on our input image than what the model spit out in the upscaling. Yes, and this is the IPA divided. I hope it has become clear what he can do. So mainly he splits us pictures that are not in the 1 to 1 format into X areas. He sets the masks for it and is then able to provide the description from this image segment per mask part. We can then hang an IPA behind it for a global description. And in principle it can also be used quite well to do an upscaling after that. Depending on how you need it or would like to have it. Especially when the output image is a portrait or something. We'll let that run through in portrait mode again. What did we get? This is our initial image. Here we see the graffiti wall again. We see the style of the Mona Lisa up here. And now the IPA is also split up here again into the game. Because he crops this image into two areas and then upscales it. And here, too, the effect is recognizable. Here we have more Mona Lisa in it, while here the model has hit very hard. Also in the background. The rough shapes are also in the new variant or in the model variant. I'll just call it that. However, from the original image we have taken the style much better in this variant than in this variant. Yes, let's see what the next video will be. I know it's a new feature on the way again. Maybe I'll cover that. Otherwise we'll probably look at the batch in the next IPA video. I don't think there's much to tell about that either. I hope to see you again. Until then, have fun trying out the IPA. See you, take care and bye!
