Key moments:
0:57 – Human Faces
4:19 – Oak Tree
6:54 – How to Install
We walk through how to use a new, highly discriminating stable diffusion img2img model variant on your local computer with a “webui” (Web UI), and actually achieve decent results.
Webui Repo: https://github.com/AUTOMATIC1111/stable-diffusion-webui
Reddit Thread: https://www.reddit.com/r/StableDiffusion/comments/xboy90/a_better_way_of_doing_img2img_by_finding_the/
A solid tutorial on installing another webui: https://www.youtube.com/watch?v=bu2xwx-6JGk&ab_channel=Aitrepreneur
***NOTE***, k installs webui from this repo: https://github.com/sd-webui/stable-diffusion-webui
you should use this repo instead: https://github.com/AUTOMATIC1111/stable-diffusion-webui to get the new version of img2img
Why are there two repos you ask???? Smart move, So did I!!! the nerds just started arguing about IP law: https://github.com/sd-webui/stable-diffusion-webui/discussions/686
Technical Support on discord: https://discord.gg/kD94kVnvE4
#stablediffusion #aiart #ai #art
Absolutely nuts
Nice one))
Good stuff
to get the closest prompt to your image you can interrogate, the clip interrogator code will spit out the closest prompt and you should have less margin of error generating photos very different from the original
Protip: Use interrogate to get the original prompt
We are heading to god mode slowly with SD
video worth the effort
can this be used with the mac m1 gpu?
instead of using a hammer, now we can use a chisel
Insane !!
Good to see you close to this, would be keen to see how you get on as you progress! Thank you!
As an artist, right autumn tree is very similar. Just compare the shapes of tree crown, it has the same shapes and structure. Of course trunk is different, but that’s is a result of higher CFG scale
Can you do a video on how to install these types of github scripts?
Thanks for the information. Very good to know! Would anyone be able to suggest a way to use the Web UI to connect to a locally installed version of Stable Diffusion but do so without being on your local network? I'm afraid I've become a bit addicted to using SD and sometimes during my morning walk at the park I'm thinking it would be great to be able to generate some images while I'm walking. I know there are collabs and there is the paid dream studio. But that gets a bit pricey for me. Since the local install is essentially free, I'd like to make the most of that (if possible) while I'm on the road.
What will SD be like when there's a normal constant download just 1 click button install , community needs to come together to make a program like blender but for stable diffusion
This wasn’t a good video
the whole tNice tutorialng but then you have a solid foundation.
It's pronounced "oiler" not "yuler "
what gpu do you have?
can i pay someone to help me set this up
how did it take you a long time to figure out the parameters? I used the parameters in the readme and got similar results to you (you also used those same parameters)
This is a good introductory video, but you kinda did do the oak tree and this feature dirty when testing stuff with that image. I managed to get great results with both autumn and winter themed prompt changes with the same image as you used. The three biggest points are:
1. Putting the CFG scale higher usually does not have a positive effect. Which value is "right" for each image varies, but usually putting it between 0.3 and 0.6 gives better results in my experience with this.
2. Higher decode step counts can help dramatically with getting you more impactful (yet more accurate to the original image) changes at lower CFG scales and also helps a lot with the "compression artifacts"
3: High sampling steps are a major contributor to the compression artifacts. Around 15-30 is usually much better in the majority of cases, preventing the majority of the artifacts!
Finally prompt engineering can help quite a lot with getting the changes that you want. Just adding or changing a single keyword isn't always enough, like what you did with summer/winter/autumn in this case. Here's a demonstration of my results using these points: https://imgur.com/a/tTrkUFb.
Hope this is helpful and doesn't come off too critical, I don't actually know the underlying code that makes any of this work any better than you do. I just want to showcase that this algorithm is more universally powerful than you show in the video!
Very useful. Thanks
Is there a google collab for this? My computer can't quite handle it otherwise.
Very cool, thank you for making this video.
I LOVE YOU
how to stop hearing yourself in headphones when recording softs?
TypeError: expected Tensor as element 0 in argument 0, but got ScheduledPromptBatch. Need help please
Thanks for all of your tips! I've just been getting into producing myself for my artists and mannnnn. It's like science lmao. Thanks a ton!
hello friend, could you upload a video of how to use this in colab?
alternative test never worked here for some reasons
TypeError: expected Tensor as element 0 in argument 0, but got ScheduledPromptBatch – anyone else getting this?
euler is pronounced "oilah" btw
I must confess that learning how to use the software is what gets procastinating to start making soft. It scares lol
Holy shit
Thanks for the video!, do you know if it is possible to run the optimized sd with this new discriminating algorithm?
I ike your videos but you do need a better microphone.
I never heard someone talking that quick.
I wish it were this simple because I have a picture of a dog. I said "a dog wearing a collar with a bell on it" and it made its ears into flowers.
v=Vke_M7N6H5k&lc=UgwjCP7W0a-KnewvlIh4AaABAg Recording is !!!
Nice thanks! (ps Euler is pronounced 'Oiler'!)
Is it possible to take a photo of a real person and make them into a 2-d flat color cartoon (not like those hideous cartoon filters for phones or PS) but actually in the style of cartoons like Hilda, Over the Garden Wall, or Adventure time. I have been experimenting with lots of different AI and all of them seem to have trouble doing this. Stable Diffusion seems to be the only one to actually attempt an artists style effectively, but there are still lots of gaps and issues.
My project requires a lot of 2-d cartoon backgrounds in styles similar to the cartoons mentioned above and I was hoping to use IMG to IMG as a way to turn real backgrounds into cartoon versions in those styles. So far the version of colab I'm working with is only up to 1.4, hope 1.5 will come soon, but do you think it's possible?
Bravo, brother!! Excellent tutorial!!!
This is not at all working for me. I am not able to regenerate the same image even with the same text and image. Not sure how it's working for you.
I already have it installed but I don't know how to do a git pull to upgrade it. Anybody know the code so I can put it in a .bat file?
I'm going to be very kind about this but in a video where you say your going to show us how to install, kinda lame you just say "following these instructions."
Euler is pronounced oiler BTW.
can u make a tutorial on how to install this exact version of sd webui in Google colab? please!
I've been studying and using Stable Diffusion UI V2 img2img and it's converting my background sketches into something good. As I'm a comic book artist, I was trying to find a way to speed up my gigs. Even with the right seeds and the right prompts, it still needs a lot of retouches. AI is not good to keep consistency, mainly for sequential art. Still, it's faster than my previous method (modeling 3D for backgrounds). Also,, it's decent to generate a character for illustrations, but when I add two characters or more, AI seems "on drugs", creating a lot of nightmarish images XD haha
So, img2img works like a charm for one character on the cover (Right now, I'm using it the create a lot of content for my advertising campaign for my products). I'm still studying it. Great for illos, not so good for sequential art. The problem is that we can't be creative when we're working for a publisher, they want consistency throughout the project. But, for personal projects, I think it's valid to experiment and find creative ways to solve it… for now. I hope AI is more advanced in the next five years.
I can't get this to work at alllll. I do exactly what you do and nada. So weird.
You can edit parameters in html with inspect tool. For example if you want more steps.
Lol, the reason your "young" man change didn't work is because that man already IS young. He is not "middle-aged" except–I guess–to someone who is really, really young. Lol
I did everything you said everything the site said and still didn’t work for me! Not only did it not add what I wanted, it ruined the picture and oversaturated it. Changed around all the settings and nothing helped
The gut isn't middle aged. That's the problem.
Was not able to replicate your results with the new AUTOMATIC1111 updates….don't know if they updated the Img2Img Alternative Script…..
It is pronounced "Oiler" not "U-Lar", the name of the famous mathematician who contributed heavily to the branchs of Maths like Graphy Theory and Topology.
Img2Img doesnt work for me, Txt2Img does, but if I use anything with an image in it I get the error – RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same – I am using a 512×512 image, any idea how to fix this error?
sadly I can't get good results with any of my own photos :/ Everything just looks awful no matter what settings I try.
presumably this won't work with the version known as waifu diffusion for the Japanese cartoons.
this tool is so exiting too use hope to see more improvements
I cant get these types of changes no matter what I do.
Nvm im an idiot, i wasnt using img 2img alternative.
I have tried this, but the test is quit a bit heavier on the gpu then the standard img2img is, not really an improvement as for me.