3 Amazing and Fun Updates - ComfyUI and A1111 WebUI!

I've got three new things for you this time a stunning new way to create high resolution images in both automatic and comfy UI some amazing new updates to IP adapter plus and a great new model to use with IP adapter as well yes hello and welcome to More nerdy rodent geekery where today I first shift my BDI towards this new highres fix method now you know when you make a high resol ution image how it starts going wrong here I'm generating a cyborg under an arch at 1024 by 2512 and um it doesn't look right at all does it that is not correct however there is a new node that I can enable up here deep shrink what happens if I enable that node and start a generation well we will find out why is it called Deep shrink well because its magic is deep in unet land and it works by shrinking noisy latent and as you can see there the generation is ever so much better let's open him up in a new tab so we can see him in his full Glory there it is okay so this is only for sdxl right now so if you try to use it with 1.5 or SSD 1B then your results May Vary in comfy UI you just have to update and you will have the Deep shrink node ready to go you can find it by searching for patch model and there it is patch model add down scale and that is the new node for automatic 1111 there is an extension there implementation of the new highis fix for automatic 11 web UI and you can install that by going over to the extensions install from URL and then pop in the URL for that GitHub repository and install it and restart as normal and here I've got the same prompt the positives and negatives in automatic obviously this hasn't got LCM so I'm just using Oiler and 20 steps again it doesn't go up quite as high as comfy I can only go as high as 2048 here but the main thing is this new hus fix tab down the bottom all I've done is I've ticked enable left all of those at the defaults and it has given me the nice New Image this of course greatly reduces the time needed to generate your images as you no longer need to do this extra highres fix step making everything much faster so that's deep shrink in a nutshell meaning you can now make high resolution images right out of the gate and it's time for the second update which is IP adapter on comfy UI don't worry automatic fans the tip after this will work in that interface too for those of you who keep your comfy UI software up toate like I do you will have noticed some cool new features on your IP adapter plus node it now supports the new safe tenses files as well as the old bin files you've got two new weight types and best of all it now supports attention masking as well meaning you can now do things a little bit like the old latent couple extension the GitHub page here tells us a little bit about these new weight types with the linear being slightly weaker than the original Channel penalty is an experimental one and there seems to be some sort of license Clash going on there so personally I'm sticking to the safe side and avoiding the use of that attention masking however is the super fun bit and this is huge the example shows what you can do which is basically assign region masks to now have both of those characters in a single image there are links for the workflows there he's got a simple and one with two masks as well so I'm going to have a look at the two mask example as this really shows visually how it works so there very quick overview you can probably see person one person two both of them in the same image all right what is going on here well I've changed the workflow ever so slightly by moving the two loaders up the top out of the way there to reduce noage plus the two IP adapter sections I have also grouped and colored so there you can see the IP adapter left and the IP adapter right and I think you should be able to see at a glance what is going on there the first IP adapter here which is the left side of the image that's got that woman in there it's generating a mask which is white on the left hand side and that is is going into the first IP adapter that's got its model going all the way through into the second one which has the rodent and just over here as you can see the mask is the opposite so it's got the white on the right hand side and then when you generate that image then you get both characters in there so we've got the rodent bodyguard and the person as well look at that excellent isn't that fun the mask is generated in this example it's a 76 8X 52 same size as the output knot as the input images so you could just make a mask in whatever paint program you like and load that in as well but you know having it generated gives you some options too do note however that the Deep shrink node doesn't play well with masking on the IP adapter unless you're using very very specific sizes it just won't work so probably best to avoid mixing those two to unless you really understand what on Earth is going on that's the second thing done and I know I'm really in love with attention masking let me know in the comments if you have any ideas for things that you could do with that feature as well remember all the stuff you could do with latent coupling yes there it is available to you now with masks the third and final update is for any interface as it's a new IP adapter face full model now it's only for stable diffusion 1.5 so there it is in the models directory we've got this new IP adapter face full and either the safe tensors or the bin file will do as you can see it's half the size of the previous one the IP adapter plus face but amazingly they've managed to fit more face inside it how did they do that the best practices from their example notebook suggests using a cropped segmented face with with a pure white background and a weight of around 0.7 as you can see there they've got their input face and uh that has indeed come out very well or you can prompt it ever so slightly differently a photo of Einstein and then you get the face in a sort of Einstein style over in comfy UI I've got that new model selected there the IP adapter face full safe tensors as well just to make sure that everything is working and I've slotted it into that first IP adapter on the left just for Giggles I've got a face which kind of matches what they say so hopefully that should come out with yes there she is there she is we got the new rodent bodyguard on the right there and she is on the left let's just zoom out so we can see the faces and I think that has turned out very nicely indeed much the same over in automatic 1111 remember it's stable diffusion 1.5 only so I've got the IP adapter selected there and the IP adapter face full model which as you can see has generated a very good likeness of her image so there you have it you can now create images in high resolution in just one step no need for that extra set of sampling you've got some amazing updates to IP adapter plus those masks look absolutely brilliant and that new face model too which you could probably use with things like my reposer workflow you have seen this video about that haven't you

Channel: Nerdy Rodent

Views: 22,709

Keywords: nerdy rodent, ai, ai art, generative ai, ipadapter, koyha hires fix, kohya hires fix, stable diffusion, add object to stable diffusion, put object in image, DeepShrink HiRes Fix, DeepShrink, ComfyUI ipadapter attention masking, Ipadapter plus attention mask, ComfyUI hires fix, New hires fix

Id: riLmjBlywcg

Length: 8min 43sec (523 seconds)

Published: Fri Nov 17 2023