ControlNet 1.1 was recently released which adds hands and faces to OpenPose capabilities. This means that the commentary in the video about needed multi-controlnet for hands is no longer accurate. I plan to cover this new functionality in a future video. Thanks for your understanding.
I cant really tell what the effects of guidance vs weight are. My assumption from looking at these charts is that when you look at (guidance 1+weight .5) vs (guidance .5+weight 1), Weight is the camera framing of the openpose skeleton (including all limbs in the shot at 1) and guidance is the positioning of those limbs (posing whatever limbs it can see correctly at 1) I came to this conclusion because I cant seem to see any difference in Guidance until those limbs are in frame, but I see a significant change to the camera framing when Weight increases. edit after watching the next section: I think you go over this, and explain its because the first generations are the most important details (what I referred to as framing) I love these videos man, appreciate it!
Thanks for the indepth research done here! I appreciate the intuition this kind of video gives versus just a "click here. put this here" kind of video!
Take the openpose skeleton, and paint away the head (shoulders and body remain only). Now it can be used to pose animals. :-) I just found this out last night. Worked for goats, horses, dogs, cats. Not so great for kittens, since those always want to sniff at the camera in the dataset.
Hi there. I just stumbled upon this video because I have a big problem combining Depth and Openpose at the same time. No matter what I try, either the pose is nothing like it should be, or worse, the image quality is butt ugly. Do yougot any solutions / workflows / tips on combining those two to ctually good looking pictures?
What I really want is a 3d posing software that will take in a 3D model and build the openpose skeleton precisely based on the camera location in the scene. The next step would be able to import the whole 3d model and have it deal with it as you pose it. Now that, would be cool.
i don't understand why this isn't working on my SDwebui. I have the extension installed, it shows up, it has its own tab that allows me to pose. But it won't give me a button to send to controlnet. Controlnet is installed. Maybe i'm missing something. if i upload a sketch of a pose into controlnet and use canny, it works and will generate a nice image based on that pose.
Hello, I have a question: At the moment, theoretically it is possible to improve images such as comic book pages, manga or manhwa to obtain better quality, greater detail, coloring, improved line art or giving style to original image panels. For example, we have hundreds of images, each containing different scenes, characters, as well as dialogue and thought captions. Using models that are not limited to a specific fragment of reality, we would be able to improve the original, add quality to it and turn a few color stains into a masterpiece. Most importantly, is something like this possible without entering commands for each generated image? Do current models cope with such huge collections, and can they perform basic tasks such as leaving captions unchanged and changing only the artistic part of the panel to a style selected by us? In short, is it currently possible to improve images to make them more detailed, dynamic and more colorful than the original? P.S. Thanks for the videos 👈(ﾟヮﾟ👈)
That depends on a lot of variables. If the original you want to improve is a crude sketch, then it can definitely be improved by SD. If it's high resolution professional artwork already then no. For automation, no, you'll need to do each manually. And you'll have to mask out text or add it back in at the end of the process because SD does not play nice with it.
Meanwhile I try to use it with a 3d posing with his hands on his thighs, generate 100 images of that same pose, I prompt the scene with descriptions, and the damn openpose gives me 100 results of a character with his hands on his ass 😢😢
ControlNet 1.1 was recently released which adds hands and faces to OpenPose capabilities.
This means that the commentary in the video about needed multi-controlnet for hands is no longer accurate. I plan to cover this new functionality in a future video. Thanks for your understanding.
Just want to let you know that this is one of the best SD channel out there and you deserve a much bigger audience!!
was hoping for a way to edit an output pose :(
I love these individual topics as most kind of just say, just do this here. thanks for this kind of content. SD NEEDS this!
I cant really tell what the effects of guidance vs weight are.
My assumption from looking at these charts is that when you look at (guidance 1+weight .5) vs (guidance .5+weight 1), Weight is the camera framing of the openpose skeleton (including all limbs in the shot at 1) and guidance is the positioning of those limbs (posing whatever limbs it can see correctly at 1)
I came to this conclusion because I cant seem to see any difference in Guidance until those limbs are in frame, but I see a significant change to the camera framing when Weight increases.
edit after watching the next section: I think you go over this, and explain its because the first generations are the most important details (what I referred to as framing)
I love these videos man, appreciate it!
Thanks for the indepth research done here! I appreciate the intuition this kind of video gives versus just a "click here. put this here" kind of video!
Looking forward to your video on 1.1
I am too. Adding hands and faces to OpenPose makes it a lot more useful and IMHO elevates it from being niche to a go-to ControlNet model
Very informative, useful information.
Take the openpose skeleton, and paint away the head (shoulders and body remain only). Now it can be used to pose animals. :-)
I just found this out last night. Worked for goats, horses, dogs, cats. Not so great for kittens, since those always want to sniff at the camera in the dataset.
Could you also give a comment on multiple subjects with controlnet
How would I go about rendering the open pose output to check my poses before committing to a length session?
Hi there. I just stumbled upon this video because I have a big problem combining Depth and Openpose at the same time. No matter what I try, either the pose is nothing like it should be, or worse, the image quality is butt ugly. Do yougot any solutions / workflows / tips on combining those two to ctually good looking pictures?
What I really want is a 3d posing software that will take in a 3D model and build the openpose skeleton precisely based on the camera location in the scene. The next step would be able to import the whole 3d model and have it deal with it as you pose it. Now that, would be cool.
so kinda like openpose editor? you can pose in 3d but not make any models, only the skeleton. also includes hands for depth maps
I do not hane send to controlnet button, but control net is installed and updated. what am i doing wrong pls?
i don't understand why this isn't working on my SDwebui. I have the extension installed, it shows up, it has its own tab that allows me to pose. But it won't give me a button to send to controlnet. Controlnet is installed.
Maybe i'm missing something.
if i upload a sketch of a pose into controlnet and use canny, it works and will generate a nice image based on that pose.
it cant detect centaurs, but if you manually edit the result, it generates centaurs well
Hello, I have a question: At the moment, theoretically it is possible to improve images such as comic book pages, manga or manhwa to obtain better quality, greater detail, coloring, improved line art or giving style to original image panels. For example, we have hundreds of images, each containing different scenes, characters, as well as dialogue and thought captions. Using models that are not limited to a specific fragment of reality, we would be able to improve the original, add quality to it and turn a few color stains into a masterpiece. Most importantly, is something like this possible without entering commands for each generated image? Do current models cope with such huge collections, and can they perform basic tasks such as leaving captions unchanged and changing only the artistic part of the panel to a style selected by us?
In short, is it currently possible to improve images to make them more detailed, dynamic and more colorful than the original?
P.S. Thanks for the videos 👈(ﾟヮﾟ👈)
That depends on a lot of variables. If the original you want to improve is a crude sketch, then it can definitely be improved by SD. If it's high resolution professional artwork already then no.
For automation, no, you'll need to do each manually. And you'll have to mask out text or add it back in at the end of the process because SD does not play nice with it.
Open pose 3d
Meanwhile I try to use it with a 3d posing with his hands on his thighs, generate 100 images of that same pose, I prompt the scene with descriptions, and the damn openpose gives me 100 results of a character with his hands on his ass 😢😢