ControlNet Guidance tutorial. Fixing hands?
ฝัง
- เผยแพร่เมื่อ 2 ต.ค. 2024
- Quick look at ControlNet's new Guidance start and Guidance end in Stable diffusion.
Maui's hands depth maps: drive.google.c...
Support me on Patreon to get access to unique perks! / sebastiankamph
Chat with me in our community discord: / discord
Control Lights in Stable Diffusion
• Control Light in AI Im...
LIVE Pose in Stable Diffusion
• LIVE Pose in Stable Di...
My workflow to Perfect Images
• Revealing my Workflow ...
ControlNet tutorial and install guide
• NEW ControlNet for Sta...
Ultimate Stable diffusion guide
• Stable diffusion tutor...
The Rise of AI Art: A Creative Revolution
• The Rise of AI Art - A...
7 Secrets to writing with ChatGPT (Don't tell your boss!)
• 7 Secrets in ChatGPT (...
Ultimate Animation guide in Stable diffusion
• Stable diffusion anima...
Dreambooth tutorial for Stable diffusion
• Dreambooth tutorial fo...
5 tricks you're not using
• Top 5 Stable diffusion...
Avoid these 7 mistakes
• Don't make these 7 mis...
How to ChatGPT. ChatGPT explained:
• How to ChatGPT? Chat G...
How to fix live render preview:
• Stable diffusion gui m...
![[UNCUT] “บอสณวัฒน์” ลากไส้!! ใครหนุนหลัง "แม่ตั๊ก ป๋าเบียร์" I คนดังนั่งเคลียร์ I 30 ก.ย. 67](http://i.ytimg.com/vi/QlD7mU2KKRk/mqdefault.jpg)
I think the quality of your hand/starting image makes all the difference here. If you can create or find a quality model at the correct angle for your picture, that's when this will shine.
I saw some one using depth map with real hand picture so it can paint correctly for shape of hands. maybe u want to try that.
Yes, I provided that in video description.
There is a way to do it with ControlNet + Inpainting I think, Aitrepreneur did manage to mix two different images by inpainting part of one image into another, if you apply that to a hand image/render/sketch it could get more consistent results.
This is what I did - I Photoshopped a good hand over the generated bad hand, then I exported just the hand and used it on Canny.
@@weizhenghuang6452 Yeah, that's definitely a valid way to do it 👍
Yes, I'm using the second method he explained in his video to have same picture first photoshoped and second using inpainting
Have you tried with inpainting? So would be like this: txt2txt > get an image > send to inpainting > painting over the hand > on ControlNet put the hand image as you did > generate
Enable Guess Mode and put 50 steps. the result will be much better
Interesting, thanks for the tip! 🌟
Cool idea, Sebastian. Did you try that with Openpose Hands?
Also i downloaded DAZ 3D to get poses, maybe that helps you 👌
Really love your videos and especially thanks for tapping into this topic! The hand/fingers problem in SD is really a big pain for me. Inpaint, controlnet, different models... all are of no avail. It is particularly difficult if the hand or finger is covering a part of the body, say the shh gesture (i.e. index finger on lips, which also happens in this video). I really wish that you could publish an ultimate guide on perfect hands and fingers fix soon! Have a good one 😉
you can do a lot by using negative prompts. and words like "beautiful", perfect hands, five fingered hands
@@fabgeb667 Unfortunately, I've tried them all but of no avail.
Click invert Input color too! you have a white background :)
1:00 I'm looking for something more like this though, keeping it almost exact but just a better hand.
Seems to me that Guidance End would primarily be used as a sort of img2img with more control. You could have very simple shapes like rectangles in the background to turn into a cityscape background that only need to be held for a short while before you let your prompt take over for more variance across batches. Nothing we couldn't do before with extra steps but now it's conveniently tucked into text2img.
Good point!
It does save some time indeed. :)
wtf. what is the point of the tutorial? there was no perfect hand till the end
o_0 bad hand was, bad hand is... time for nothing
If they somehow figure it out to lock parts of the prompt to the region handled by a specific ControlNet instance (or specify separate prompt for each ControlNet section), that would probably again bump up the quality a notch.
Use gues mode and do 4 iterations, instead of adjusting values
Interesting, good tip!
in ControlNet, how do you get additional tabs for control models like he has in the video? i only have one
See my video on Multi-controlnet. th-cam.com/video/cNIHZInV3mg/w-d-xo.html
File "C:\Users\lenovo\AppData\Roaming\Python\Python311\site-packages\torch\cuda\__init__.py", line 284, in _lazy_init
raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled
Photopea is awesome. Its developer is awesome! Wish there were more devs like them.
Agreed! Now I do mostly use Photoshop, but since all the shortcuts are the same, I use Photopea for videos so everyone can join in 🌟
Great video like always. It would even be better we could use this during inpainting somehow as i always want to change a specific thing in my images.
Maybe also for next time you could change cloth or hair but keeping original input
Helpful :) I will try to add rain drops
If you look at the image, you can still "see" the original hand in the modifications, which suggests that the "idea" of the hand's position is forming pretty early in the process compared to when you're introducing the modification. I think you're likely to have better success if you adjust the start time earlier until it no longer seems like the new hand is a modification of the old hand.
I think the only way hands will get fixed is a model or LORA specifically made to fix hands by inpainting.
It seems like using depth maps with skeleton from the beginning is decently consistent. Using either that blender rig or now there's an extension with a library of hand depth maps, plus can use openpose editor or the new 3d skeleton editor to get the main pose. Part of the issue in this particular case was trying to change the hand of an existing render without changing everything else.
For some reason the resolution of the controlnet hand **inside Inpaint** is way off and I can't get it to work anymore. An update must have broken it.
It still works in text2img though.
I'm not sure but maybe guidance end helps reducing producing too much details for eg background elements similar to the weighting described in this video: th-cam.com/video/HkLUmTJoyhw/w-d-xo.html
lol I just want to point out 'peace sign 2' in the hands pack is an offensive jesture in the UK.
We shouldn't even have to deal with hand problems there's got to be a way that they can fix in the software.
am i the only one who's controlnet has been broken since 1.1 update?
Did you see the LoRA thing? I used some ready LoRA models successfully, but couldn't train one myself. Would love to see a video review of it from you!
which LoRA models did you use?
@@JohnToridas Various character LoRA.
Shouldn't the "invert color" box be checked? I mean it literally says "Invert colors if your image has a white background"? Quite confusing if this isn't actually necessary.
You would think so, and I've tried it and can't make sense out of it. Sometimes it works, sometimes it's in reverse. Either I'm doing something wrong or there's a bug somewhere.
thats one frustrating thing about AI rendering...hands!
2:13 💀
Scribble is shit for hands. Use depth or canny.
on tiktok i see videos where art is moving and has many fotos, i want to learn it but i dont know how its called or what technique it is. the tiktok creator is not responding aswell 😅😮💨
Deforum. Check my animation tutorials.
Its definetly a interesting feature and I will try it out but so far I feel that fixing hands with inpaint might be better at least when you dont want a specific hand gesture.
Inpaint + ControlNet probably is the way to go
I saw "can fix hands" -> i click xD
chatGPT jokes 🤣🤣🤣🤣
which GPU is he using? is blazing fast
Rtx 3080
So now we will logically finally be able to create real star wars lightsaber fight or knight fight with proper handle/hands etc
I like your videos! I can’t seem to install controller properly on my Mac. I see it there and openpose as well but there’s no send to controller button under the stick man in openpose
you talk so slow bro lol
this will be really interesting for animations, thanks for sharing!
Thanks for this share. We're certain this will only get better and we will have perfect hand control sometime this year
this man jokes always got me 🤣
I see you're a fellow dad enjoyer 🌟
Great tip man, i was wondering yesterday about it.
why is your control net UI ControlNet - 0 and ControlNet - 1 next to each other instead of below?
They keep changing the layout with new updates
⭐🙌
So in theory we can fix the text or adding UI text in SD using this feature right ?
I did a video on text workflow, check that.
Thanks.
Magic poser has hand presets you can use
Good tip!
It is so fast… with my amd it would took me 5-10 mimutes
Vroom! 😂
@@sebastiankamph sry, don't rly understand this ^^" - but still a great video
Finally. Hands!
As the kids say nowadays... first comment
As the kids say nowadays... fail