NEW Outpaint for ControlNET - Inpaint_only + Lama is EPIC!!!! A1111 + Vlad Diffusion
ฝัง
- เผยแพร่เมื่อ 29 ก.ย. 2024
- The new outpainting for ControlNET is amazing! This uses the new inpaint_only + Lama Method in ControlNET for A1111 and Vlad Diffusion. The method is very easy to use. In this Outpainting Tutorial I show you all the settings you need and also my img2img method that gives better results
#### Links from my Video ####
Create ADs in A1111 • Make AI Ads in Flair.A...
Map Bashing with ControlNET • Map Bashing - NEW Tech...
#### Join and Support me ####
Buy me a Coffee: www.buymeacoff...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord
![How to change ANYTHING you want in an image with INPAINT ANYTHING+ControlNet A1111 [Tutorial Part2]](/img/n.gif)
#### Links from my Video ####
Create ADs in A1111 th-cam.com/video/LBTAT5WhFko/w-d-xo.html
Map Bashing with ControlNET th-cam.com/video/z6Xwh9G24uw/w-d-xo.html
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotutorials
Join my Facebook Group: facebook.com/groups/theairevolution
Joint my Discord Group: discord.gg/XKAk7GUzAW
Intrested extensions Face swap
My controlnet version is stuck at v1.1.200, checking for updates say it's the latest one, please help.
i dont see any model near inpaint only box ,help please
@@boostergold4101 did you do the update? What does the version number say? You can also remove controlnet and reinstall it. But make sure to keep the models or you need to download all of them again
is it possible to download the models manually? ((theres one on civitai by the ally, works but not sure if its the same one))
FYI, if you change both the height and the width it's going to stretch your image. If you truly want to outpaint from 512x512 to 768x1280, you should do it in 2 steps. So first go to 512x1280, then repeat going from that to 768x1280. Hope this makes sense. Cheers!
True. In the example in the video his base image was already 768px, so he just adjusted that upward to match it, then only changed the width...but helpful tip overall!
Yeah, I noticed that on the first try. I realized that it paints wherever is the difference in ratios as opposed to initial scale.
mine just keeps filling the canvas with the existing image. im not sure why
@@shabadooshabadoo4918 low denoising setting
Hmm... All I get are brown bars at the top and bottom of the image. I'm trying to make it taller.
where do i get the link to download the inpainting model?
Hello, does this method work on SDXL 1.0?
This is for sd1.5 not for sdxl.
Why is everything in SD so clumsy, lossy and timeconsuming... I am just turned off to even spend time learning the new features anymore as I feel all this will be hopelessly outdated and silly in like 6 months time.
Hmm, I updated the ControlNet extension, but the control_v11p_sd15_inpaint sadly doesnt show up in the model drop down :/
Same same, Do we just download it directly from Hugging Face?
@@Maltebyte2 yeah i just googled the name and found a download, probably what you said
Can you tell me what directory i have to put this file? Thanks@@DasHnezzEdits
\stable-diffusion-portable-main\models\ControlNet @@Maltebyte2
Help. I cant get the model controlnet v11p inpaint... It's not in the selection of models. Since you said it'll get downloaded automatically the first time I use llama so I left the model to "none". But for some reason mine wont. It's still not in the selection even after restarting my terminal and webui. Am I doing something wrong? thank you
Having the same issue, shows as empty. Were you able to figure it out?
@@FingerThatO I forgot already the exact step by step but as far as I remember I was able to fix it. Downloading the model manually. I probably asked that on reddit then somebody told me to download the model and they sent a link. I'll send it here if I can still find it
@@vonpheusarts6948 i have downloaded it too, but which folder should i put it in
This doesnt seem to work anymore in Automatic 1111 1.6. I just get black bars...
Note: You'll want to use the exact same checkpoint for outpainting as was used for creating the image in the first place.
I did some testing and even rather similar models perform very poorly, mostly resulting in largely blurry messes.
I can get control_v11p_sd15_inpaint model, it didn't instaled automatically.
the model doesn't download for me, i get a "You have not selected any ControlNet Model."
Yeah, this is frustrating, the inpaint+lama shows up on the first dropdown, but the control_v11p_sd15-inpaint doesn't show up in the model dropdown. I verified that the model is in my models folder. What Am I doing wrong?
yes, do you have fixed it ? im running the forge version
Where do you get the controlnet model if it doesn't auto-download?
same issue here
No, it does not automatically download the Control Net model on my first use.
It didnt automatically download the model "control_v11p_sd15_inpaint" when I tried this... how do you manually download / where does it get saved to?
did the update and got the inpaint lama but the control model does not download no matter what I do. Is there a manual way?
@2:35 my RESIZE MODE buttons are present, but they are greyed out, I cannot click any of them, it is set to CROP AND RESIZE and cannot be changed. Ideas? (I'm using Vladmandic)
Resize mode is greyed out as if you don’t provide controlnet image
I have disabled section "Resize Mode" on img2img tab. Could you please tell me why does it happen?
I have disabled all of three resizing options there.
UPD: but it's available on text2img tab, also in the ControlNet section.
Resize mode is greyed out as if you don’t provide controlnet image
@@SynthArt_Studio You are almost right. Indeed, the problem was that I didn't provide the image but I thought it was set! And it was set really there. But not as I thought :) ControlNet automatically placed my image into controlNet's image field and I thought I had already set but it's not :) When I set it manually by dragging it from my explore - then all worked fine.
In other words my image was there but it wasn't set 😀
as others said, it does not download the inpaint model
How would you offset the image so that we can extend, for example, only to the right of the original image?
same question
You can click the edit button in the image field, same place you’d normally go to inpaint in automatic, and then extend the border to whatever size you want. Then you can click the protractor icon to update the resolution
I was about halfway through the video when it disappeared!
Just getting an error instead of the inpainting models download automatically
How does the download of the model trigger? I have not managed to do it
Hello, you mentioned that after selecting the processor inpaint the model will be downloaded automatically but to me it doesn't work. How to get the model downloaded? Not sure why but my controlnet has only the Openpose and Canny model while it has all the processors. How to fix?
So for 1. it does not automatically the inpaint model (and I can also not find a download anywhere) and 2. my main image also changes, it doesn't filly only the empty space.
what if I want to keep the original character in the original image, and just want to outpaint the image without affecting the original character?
i got this error for lama: "ModuleNotFoundError: No module named 'annotator.lama'"
I hope they'll soon implement a method to outpaint only to a selected side, like only to the right and not always equally to both sides. Or let's hope that invokeAI catches up some day and implements ControlNet, since their Unified Canvas is the most intuitive and user-friendly way of out- and inpainting. However, as usual great video Olivio! :)
You can use the poor man outpaint script to expand image.
hello help. I cant get the model controlnet v11p inpaint... It's not in the selection of models. Since he said it'll get downloaded automatically the first time I use llama so I left the model to "none". But for some reason mine wont. It's still not in the selection even after restarting my terminal and webui. Am I doing something wrong? thank you
Thank you. I wanted to try it, but SD does not automatically download the model. How else can I get the model?
I just get the error "RuntimeError: Given groups=1, weight of size [320, 9, 3, 3], expected input[1, 192, 128, 175] to have 9 channels, but got 192 channels instead" when using anu SD 1.5 model or sampler. anyone found a fix for this?
I got this too and I think it's because we used an XL checkpoint.
I don't think the ControlNet model is downloaded automatically. Definitely doesn't do so for me. Models not installed don't show up in the list to begin with 🤔
Also, I get this is a quick short video, but it left me wondering if this only works well when using generated images and sticking to the original prompt and model. Here's hoping for a more detailed/varied video 😉
It didn't download automatically for me either, do you know where to get it?
@@dosexu1 one on civitai by theally works
same issue, module is not present, download it, nut still won't show and won't work for me :(
@@tizianomayles3084 it did work for me after downloading the model. I downloaded what I think is the official pth model (by lllyasviel, from huggingface)
@@tizianomayles3084 Yeah that works, i downloaded it from huggingface as well
I REALLY wanna get into this, very excited, but got ZERO idea on the basics.
Where can I learn? How is it all called? Do I donwload a program? (My PC's jurassic.)
I don't get it, he sets up a higher resolution, but the main image doesn't change even one pixel, as if it was masked, whenever I use img2img the base image will change slightly. Is there something I'm missing? is there a mask somewhere?
ControlNet is the most powerful tool that we have in SD.
NOT working anymore , same workflow , you'll get an whole new generated picture that overwrite your starting image, I don't know if it is due to controlnet update
if you get that result, you are doing something wrong. did you enable the controlnet?
@@OlivioSarikas I am sure I have the right setting ,because 3 days ago it worked fine ,maybe because I updated my controlnet , when I try to outpaint a face, it gives me the whole new generated picture without my original input, when you outpaint an object,it is not easy to tell, but when you outpaint a face ,you can tell the difference .
I mean it replaced my original picture with very similar picture, but it's not the original one, especially obvious when you try to outpain a face,
So with this method it seems the outpaint always expand outwards from the center of the image. How should I approach an outpaint with the original image offset? for example only outpaint the rightside
Thanks for this latest update Olivio! With the download step, the image will usually already exist within the outputs/txt2img-images folder/directory so can always be uploaded from there, unless wanting to create a copy as part of the workflow. Recent updates of the Infinite Image Browsing extension (highly recommended) also has a context menu option to send an image directly to the t2i and i2i control net panels, so you can just click the tab for images, right click the desired image and send it straight to CN.
You've just sold me on that extension. It always seemed crazy to have to shuffle an image you're already using around so much just to put it in another tab in the same window.
If I'm not completely mistaken, if you just don't select an image in controlnet at all it will just use the img2img one anyway so that's even easier.
@Janek Nice. Just tested generating a cat, sending to i2i and enabling canny. Changed the cat style in the prompt and the cat changed but the image structure was nicely coherent.
Thanks this is amazing! Btw you don't need to download the image you can just copy/past it, it works!
Do you know if it's possible to pick a direction to outpaint? Can we outpaint only on the left of the image, and not on both sides?
Intrested extensions Face swap
does this work for real world non-ai generated photos too?
Hello anyone knows how to instal the model for inpaint+llama (control_v11p_sd15_inpaint ) no showing on models
can we use controlnet to restore color in retro photos ??? - if yes = this is the idea for one of your next vids :D
you should use (:1.2) instead of (()) it's more convenient especially if you go to high strenght or lower than 1, use shift + arrow up or down to change the str
a life saver! thanks for making this tutorial! i finally managed to get outpainting to work! do you have any idea on how to extend in one direction e.g to the left only or upwards? thanks again!
Thank you very much 😍
hey, have you found an answer? :)
SahiNama covered this on his channel too, if you’d like another tutorial on it.
Can you change outpainting to only increse the height of the image to the bottom and treat the image as the top of the image?
bro, you look like you're on holiday every day 😀
It sort of works, but it also changes my original input image, not just the outpainted parts?
Am I doing something wrong?
If I use img2img tab, it changes my input image, but does correctly resize and fill.
If I uise "Inpaint" tab with exactly same settings, it does not outpaint but just resize (giving squished image)
It's amost as if it won't invoke ControlNet at all. (Yes, it's enabled and works for other things)
EDIT: Ah. Checked my output. I am using an SDXL model anf that doesn't work with the control_v11p_sd15_inpaint model.
I need some SDXL control net inpaint model it seems. I just haven't found any that works.
what about using outpainting to extend images that were not originally generated in SD? my results show it struggling to match the original artstyle
control_v11p_15_inpaint is not downloading for me... How do i get this model then?
I'm having this same problem. Were you able to resolve this issue? The resize and fill option doesn't appear for me either.
@@gaviera4657 i didn't sorry
Thank for video, but you didn't mention which positives and negatives you are using during inpainting mode of ControlNet. Or are they not important? Or use as in script Outpainting mk2?
Does this still work? I'm following this guide and getting this error: ValueError: No mask detected for ControlNet inpaint
The outpainting algorithm is decent in guessing what the extra data on the sides should be. However, I prefer the extension openOutpaint, since you can control which side is extended and by what amount of pixels.
Olivio, is there a comfyui method, that gives the same awesome result???))))
For trying to get different poses and now for trying to get out-painting,. ControlNet is not working, I repeatedly get this error message: **AttributeError: 'ControlNet' object has no attribute 'label_emb'** It refuses to begin to render, only pausing for about a minute then returning with the error. VERY annoying. It fails with other checkpoints. I've set my SDXL to update on launch and have checked for updates for all my installed extensions. No luck whatsoever. Any advice?
Great ! Have you tried it out with other type of images? Like different artist's styles?
Thanks !
Inpaint_only + Lama for me was worse than inpaint_only controlnet outpaint especially bad for real pictures
Thanks for the video. My machine can only handle one task at a time so I get to try it and see one go at a time. I did wonder about the prompt. Does the prompt stay the same as when I made the image, or do I change it for what I want the fill to look like? And can I use CN with maps to populate the wider area? I'll give it all a go and find out, but that's my question. I do wish I could specify where in the new image my original image was positioned. Centred every time is a bit boring. I guess I just render two 512 and stick them together. Trick would be matching up their sides. We'll look back on all this and none of it will be as fiddly in a year or so.
RuntimeError: You have not selected any ControlNet Model.
When I try this the original image and subject also changes, it doesn't just outpaint the original image. Am I doing something wrong?
I tried this method to finish a character whose feet were not visible, but I couldn't get any results... does this method work for this kind of outpainting?
Hmm, I've tried this a few times, but the transitions always remain clearly visible. It's not pleasant. :(
the model don't download where can i find it ?
this workflow don't work in ControlNet v1.1.230
hmm, why is that I follow this step, I'm getting a stretched image instead of an extended image? I feel like I may have forgotten something.
Enable resize and fill option.
yes, resize and fill
@@OlivioSarikas I have resize and fill as well. I don't know why its not working lol. It keeps stretching. I sent image to controlNet, then added widgth.... Inpaint Lama, control_V11p_sd15_inpaint, ControlNet is more important, Resize and Fill. Denoise .7 I played around with it alot but still no luck. Hmmm
@@narva5440 Yeah I did that lol. I think i did everything that was in the video. I'm pretty sure. I enabled resize and fill both on top and the bottom where ControlNet is. Not sure why its not extending.
@@AndyHTu did you tick enable controlnet? I often forget that :)
Thx. guys
I don't know why, but when I do it, the program changes my whole image
I'm a bit new to this, so help me understand something. Instead of going through all the work to expand the image with outpainting, why not just generate the image at the size you want to begin with?
Why do I have very visible borders when expanding the image, all the settings are as in the video.
thanks for this, I just tried it, ist there a way to blur the edges? I have pretty rough hard edges in my results
This stuff is totally wicked!
Confusing and lacking of necessary info, now I gotta go and watch another tutorial and hope to find what this was lacking.
How about enlightining me about ControlNet 2.1, I cant get it to work.
It doesn't work most of the time, it gives me unrelated colors in the filled area.
Everytime I generate an image I get people in either side of my main image. Is there a way to stop people from appearing in the background?
has this method changed? i used to use this but i can't get it to work now. I haven't created images for a few months.
Well, Its giving garbled result to be for the areas to be inpainted. Any recommendations on what model to try for this purpose? I feel it could be model dependant as well. I tried a regular photograph with 512x512 dims and was using your method to enhace it upto 1200x512. The central part of the image as fine as is but the left and right side were off. Results didnt generate anything similar to the source image.
Same, I am wondering what's happening, I followed every step of the video.
I think i found the reason, its the denoising strength, if you try to keep it low like 0,5 it will be just colored lines, you have to go at least 0,6 or 0,7
Great stuff! Basically what Adobe are doing with Photoshop Generative Fill, but for free and without content filter. 😉
There any tutorials in populating an image with countless people?
error 'NoneType' object has no attribute 'rsplit'
on DirectML cant use it atm. xd
Damned annoying - getting pretty sick of chosing the wrong card for SD - who knew I'd have been better with less ram but nVidia
Supper helpful!! is it work with videos??
Really digging the new videos! The practical workflows are just great.
thank you very much
Can you send me an API of this workflow?
what happened to dalle3 outpainting. it was so easy. sheesh.
Это тот самый француз , которого так ненавидит Хачатур?)
Amazing content by the way how are the results with real photos or photorealistic images, also which of your tutorials you recommend following to have the installation of Stable Difusion to make this kind of outpainting?
Am I the only one getting darker stripes where the new outpainting is?
Hey thanks for your great work. I followed this tutorial, but the issue I am seeing is that when the outpainting is done, the original "internal" image has changed as well. Your examples seem to show that the original image and the outpainted image, the internal image stayed the same.
Mind you I am also getting the "RuntimeError: The size of tensor a (3) must match the size of tensor b (0) at non-singleton dimension 2" type of error, so it may be in issue with size mismatch somewhere.
followed too the T its just resizing image
ahah a guy a follow put this already online :D y are late olivio :D ^^
You could even delete your main prompt to allow it to imagine even more potentials, can also use masks for doing only small or particular areas.
interesting. i have to try that
Very nice. It gives me great results practically on the first try. One question: Is there a way to outpaint in only one direction, for example only left instead of both sides at the same time?
well the "old" outpaint did exactly that (script "poor man's outpainting") you can still use that instead of control net, results should be about the same, at least for me
@@thFaust From what I remember that old outpaint method worked really bad for me but I'll have to try it again to compare with the ControlNet method.
@@Macieks300 yeah, do that. I use inpanting models with the old method and denoise of 0.8, results are similar to the video
What if I want to extend the image only into a certain direction?
Incredible, thank you. I think outpainting is finally viable. If people are finding seams where the original image meets the new parts, you might want to place the oupainted image back into img2img and then generate a new image with a very low denoise strength. That should fix it.
hello help. I cant get the model controlnet v11p inpaint... It's not in the selection of models. Since he said it'll get downloaded automatically the first time I use llama so I left the model to "none". But for some reason mine wont. It's still not in the selection even after restarting my terminal and webui. Am I doing something wrong? thank you
@@vonpheusarts6948 Yes, I had this same problem as well. It did not automatically download, just find the page for controlnet for hugging face and download it from there, wamn
Yes, thats an issue Im having with this method, the seams are very obvious. Will try that trick. Thanks
Can't wait to start this method with all my previous pictures and see what I get.
Hello, thanks for the video, however there is something i dont get :
I feel i understand every settings you are using but when i generate, it makes a very similar image but the middle (so the base image) is also re-generated. So i find myself with a very similar image, extanded.... but the middle, the base material, changes.... any idea why ? Thank you and keep going with the videos =)
do you have controlnet enabled? is your denoise setting high enough? did you load the image into controlnet?
i can't use controlnet at all even though i have medvram, xformers selected. It does 1 generation and then after that CUDA out of memory. i have 8gb ram. How can i make this work? i also select lowvram on controlnet. Can anyone please help?
You have a problem somewhere because I have a RTX3070 so 8GB too and I can generate HD images with 3 ControlNets fine :/
Do you update your A1111 often to get latest additions ?