I learned that the AI looks for a solid reference to make the hands correctly, and that little bone that we have before the hand, This tells the stable diffusion what orientation it should take. 😎 Without that bone he is like Adam on Mother's Day. 😆😅😅😅😅
When the AI does not do what it is told, put that order as close to the first as possible and then if it does not pay attention, indicate 1.4 so that it obeys. It is possible that your data store does not have the requested reference, perhaps it is in another name, such as: sphere.
I have been clicking around on tutorials and guides to improve my results with SD and yours are by far the best I've seen so far. Concise with no fat or filler, comprehensive showing different ways of tackling problems, and you do a great job of breaking down complex concepts into steps that even a complete beginner can understand. Keep up the great work!
I learned a small fraction of these techniques a few months ago and then promptly forgot them through lack of use. This video seems to be PERFECTION for creating the best hands right now. Hopefully the state of generating hands wil improve but THANK YOU so much for your efforts, not least your webcam tool.
You really go above and beyond to create the most comprehensive video possible. Between your last video and this video, I switched to ComfyUI since A1111 just broke one day and wouldn't generate anything other than noise (or just give the nan error). I thought I wouldn't be able to follow your tutorials anymore, but there's so much info here that will even help me in ComfyUI. Thanks for these super detailed videos and please keep up the good work. Also wishing you a speedy recovery, take care.
Thank you omegasuperveryfking much for the kind words. With comments like this it's so much easier to create content. Hope it helps and, maybe we could try to see why A111 was crashing. Have you re-installed? I'll try to keep providing value for as long as I can!
@@Not4Talent_AI I did try reinstalling it but it didn't work. I also made a reddit post about it asking for help. (I initially put the link here but then TH-cam kept deleting my reply for the links so had to delete it) After reinstalling, things became worse for some reason. As the one or two generations I was getting before were also replaced by nan errors. I changed my nvidia drivers to studio drivers instead of game ready drivers to see if that helps but it didn't (yes I performed a clean installation). To be honest, I think it's mostly because my laptop is potato (4GB VRAM) considering image generation standards, but I'm just confused as to why it worked before. After switching to ComfyUI, I do miss all the extensions from A1111 but it works. There's just more I have to do to get the work done. Also, ComfyUI uses a lot of CPU RAM in lowvram mode so I think that's the reason it works on my laptop. From what I've understood, it offloads the model onto CPU RAM whenever they're not in use. It also uses CPU for latent generation (maybe? not sure. read it on reddit). As a result I can generate a 512x512 image in about 30-40 seconds compared to A1111 which took about 2-3 minutes.I still wish I could use it though, cuz it makes some things way easier, so a combination of both would be my preferred way. Sorry for the long reply 😅. Hope I didn't take too much of your time. Please just continue to do what you do. Your videos are more than helpful no matter what I use since a lot of what you tell are fundamentals of image generation. I look forward to your future content!
By far the best SD creator!! I check for your content daily, keep it up, nobody makes content on your level for people who are familiar with the tools SD has to offer.
Thanks a lot, really! Tbh this video is a lot of manual work, but I hope I'll find better ways as things improve. but thanks again for the support, means a lot
I just discovered your channel and I really like your tutorials so far, thanks and I subbed :) But what's really refreshing is to see an artist embracing generative AI instead of foaming from the mouth about it. Really shows what a great tool Stable Diffusion can be for artists. The rest of the tutorial channels I'm subbed to are mostly tech geeks who've been using SD since its release and have gotten a lot of experience this way, so yours was a nice change :D
Thank you very much!!! I will start now that Controlnet Is finally working on there, even though I've been avoiding it cuz my gpu doesnt like it very much xD
I encourage people to go into any illustration software, and just do back and forth editing between manual edits to guide the generation explicitely and prompting with controlnet. You can get amazing results with just these 2 simple steps and some training! The more you do it, the better your images will get and so will your artistic skills (and it's super rewarding when it finally does what you want! haha)
Oh yes hands.. The Archenemy of Stable Diffusion. For me it also got better with SDXL but still it's my createst enemy.. Thanks for this helpful video 😊
True, I didnt mention it... sht... Sdxl is better with hands, so as controlnet gets better for it, it will be easier to make them there. Even though atm I couldnt really get good generations using CN and SDXL so I left it out. Still should've mentioned😂😂 Thank you!
@@Not4Talent_AI Well for my last model DEMON CORE I needed to create a picture with a woman holding a metal ball. It did it well but I needed about 10 trys to get it perfect without using inpainting. There were mostly small errors like holding the ball from or a strange position of the finger. Overall it did quite well ;-)
the goat is backkkkkkkkk, epic video, p.s. for my workflow I send images to the photopea extension instead of using inpaint sketch to be able to fully edit the canvas easier.
Thanks soo much @Not4Talent, you got me giga interested in Stable Diffusion, and I'm seeing great progress soo far! you're not just a great teacher you're also entertaining to watch
Patience = Power 😉 this is why I *love* manual work. More accuracy, control and quality. Now that I look at it, it's not really *that much* more work, looks totally fine. If I were to try prompt-only with no cnet and tweaks, I'd lose all my hair after the first session of a couple hundred images But I get it, people just want to get professional looking jackpot image in minutes ¯\_(ツ)_/¯ this is why we will stay on top of the quality work
tyty! There are some faster ways to do this with just inpainting and a bit of luck now. Mainly with anime and Ponyxl. Or using SDXL with models that are nicely trained. Might want to check that too if you struggle with some hands research probably bout 1 week or so, practicing a bit more
Very cool techniques! I noticed that your videos pack in a ton of info at lightning speed. It would be awesome if you could do some 'part 2' versions of videos like for this one where you do one or maybe two examples of the technique from start to finish. This could be good for a live stream that gets saved for later viewing too. So you wouldn't even need to edit the video which I'm sure takes a ton time.
thank you!! I've been thinking of a way to do that but I still havent figured it out. I'd like to keep the video's on a steady standard when it comes to editing and style, so mixing it with long unedited format videos feels weird. Also, I would stream but my internet is bad af... So yeah, I really agree with what you say but I have no idea how to go about it xD The suggestion is pretty nice tho, I have to think about it more. Tyty!!
Dude why are you so underrated?? I have watched tons of AI content creators and tutorials and no one explains it with so much passion and accuracy. Truly game changing. I have a question tho: What can I do if I need the same hand, but on other characters and in different positions? I guess the answer is "train a LORA" but is there any other way? If not, can I train a LORA just on hands? I also need to train the LORA on both hands.
thank you so much!! I'm not sure I understand the question tho, do you want to use the same hand pose on multiple characters? Is it in the same position always or does it change position but not pose? You can train hand LORAs, like this: civitai.com/models/99237/hand-on-own-chin-or-pose (it doesnt give amazing results, but it's very fast and makes a base image that's easier to clean up than just having to star all over again for a new character) If the position doesnt change, just the character, I think hands should be similar enough where you can just change skin tone and reuse hands. If position and character change, I think it depends on what the goal is, but You could create a library of openpose + depth + edge references to re-use and pose properly with something like the "depth library extension". then just remember which weights work best for the checkpoint you are using Hope this helps!
I guess in the time that all the manual work takes, you could just have created 300 other images more or 300 variations of the one you like and check if one of these got the hands right. ;-) I've also the impression that the bad-hands5 embedding makes it at least a little better. I saw it at least trying to fix fingers multiple times. But let's say you finally got SD to get the hands right. Now you want to upscale the image and in that process it has to be created again, that is if you actually want new details and not an image that is just a lot bigger. Now you can choose between a blurry image that is forced by c-nets to be as close to the original as possible or a nice new detailed, similiar looking image that got the hands wrong again... after minutes of computing. Maybe it's just me, but there are days, where I want to have nothing to do with this stuff anymore. ;-)
If you have the computing power then yeah, maybe you can do that. But I think having the manual tools to fix stuff is straight up more consistent. I agree that upscaling can be tough, even though not as big a problem as you make it look like. (imo). And yeah, AI is not as easy as it looked in the beggning, at least not if you want actually good images that look professional. Which is why I'm a huge advocate for manual work, since it lets you learn and overcome this issues while AI advances
9:50 Little pitfall here: If you use inpaint sketch or Inpaint for small things / small masks, you need to reduce the "Mask blur", to not lose your change in the too large blurred area. So to get those thin fingers inpainted, you would need to lower it quite a bit to actually get them generated, depending on the resolution of the image. If it's 512x512, you would probably have to lower it to below 4 here to actually get a finger. In 1k x 1k you would maybe need 8 or even less. I regularely miss this and get frustrated before it plops up in my mind again that my blur is too high.
as an artist myself I tempted to use SD like this. but i think it's more like using SD as replacement for pinterest. openpose is good. I definitely will explore that. and ultimately train my own characters later.
4:00 oh yeah there is Paint toll sai. it has some really good Linework for vector use. you can trace over it and even edit and add pressure to the lines. if you only have mouse. I feel like you know what I'm talking about.
You're so good with Photoshop, I understand you need AI to create 60% of the image then, the AI will hold your beer while you do your magical tricks. Your workflow is super interesting for lamers like me. On my side, I just generate zillions of images with diff prompts technics, Lora's, checkpoints and TI (even if I know they make hands worse usually) then throw it to img2img and hope I'll get the jackpot without be like "inpainting" anything.... and of course, each image is saved on my side, just in case of, "you know", lol... throbbing the perfect ticket. I guess, it's a two school vision : yours, you'll have to put efforts learning something else, dominate your tools, dig a stone and fine tune it... mine, just blow it up and sure, you'll be surprized, at some point but you can't be sure you can reproduce it... and it was a lucky day and the lucky days are not common... shrug
Thank you for that compliment ahhahhah. I have been using it for a long time but, tbh, I think most people can get to a level where you can do this in about 1 or 2 weeks. So if you feel like it, I'd encourage you to try it if it is something you will need. If not, your approach works too xD Iterate until the right answer comes up ahhaahha. Ngl I do that too whenever I'm not focused on creating something and just want to generate stuff on the side. hope that, either path you choose, you accomplish what you wanted! Thanks for the kind words too :3
I only work with Midjourney but anyway your videos are one of the best sources of inspiration to solve problems related to posing and composition I've ever found. Thanks, man.
Can you share what exactly prompt, denoising strength and cfg scale you've used while inpainting sketch? I'm only beginning to learn SD so i dont have idea about control net, but your results in inpaint sketch were pretty cool so i want to achieve the same.
oh, I changed pc so I dont really have the images. But I usually use different ones depending on how accurate the sketch was. IF I used a good color and a nice shape, I'll probably keep it with a mask blur of 1 or 2 and a denoise of 0.2 to 0.35 If I did a bad sketch that was more of a "base" for SD to iterate from. Then I'll use a higher blur and higher denoise. Like 0.4-0-75 CFG I usually just move between 5 and 8. Mainly 7 . Prompt you always prompt for what you want to see inside the mask, unless you also want to give some context about the rest of the image.
I wonder if you could make an extension that used a GAN (adversary network) to regenerate the inpainting of a hand until it looks like a source image, so you could just walk away and let the software to the generating until something good comes out....
Hi there. Great video, although the answers and methods are great. But that's only a single person. It sadly could not help me with MY specific problem. I am trying to use CN will all its possibilities, mainly OpenPose, Depthmaps and Canny to generate images with dynamic poses with TWO (or more persons) closely interacting with each other. For example as "simple" as a couple dancing. But I can't get any results, even with your tips here on different CN models. I also used PoseMyArt from the beginning. But whatever I tried in the last couple of weeks, months even: Either, the result is butt-ugly or SD (or rather CN?) simply does not what it is asked for. If you know how do get something like this done. PLEASE, PLEASE do a tutorial on that.
Hi!! thanks! that's literally the next video. If you have a specific image you want to see done, please mail me a sketch or something and I'll see if I can add it as an example for the vid. I'm currently scripting so I think I have most of the workflow layed out
Thanks but no thanks. I wouldn't have the patience... I think I will make them wear mittens instead. I run out of memory when I use two controlnets anyways. but thanks for sharing your experimentations.
@@Not4Talent_AI hopefully a miracle solution will pop up and you'll tell us about it. Some whiz could come up with a script like LH(50,0,100,100,100) -> Left hand thumb bent 50%, index straight (0% bent), other 3 fingers bent 100%. Idk...
The problem with this is that this is actually working for your picture and then we wouldn't be fulfilling all of the stereotypes that no-ai protestors are spouting. (Seriously though - thanks for the compilation of techniques. I knew most of these already, but I somehow never came across the idea of just painting my own canny controlnet...)
After drawing the hand in photoshop (11:03), which pre-processor should I use? I found it a bit confusing. Should the background always be black? I tried to do it but it was messy, the edges of the hand went into the image, it was strange. I'm using the "depth map library" to position the hand on the image, so I don't know if there's a problem.
hi! none should be the preprocessor. So literaly no preprocessor should be used if the lineart is done properly. I guess using the hand library extension is fine to use btw. If with "none" it still fails. Maybe try making sure that it is properly position with another editor like photopea or photoshop
@@Not4Talent_AI I finally succeeded, but I had to use a high denoise strength of 0.75. Last question, is it possible to draw "perfect" nails with this? Thank you.
There are some pretty big mistakes even in the thumbnail, but with some more perspective training and study from life you'll be perfect at hands! Keep it up ❤
😂😂😂 I mean fixing is in photoshop. Hands are in SD. But yeah there is a lot of manual work. better cn models and optios have been released recently though. Now there is not that much of a need for fixing. Even if still a bit
havent tried tbh, sounds tough. Maybe prompting with inpaint sketch (trying to help aI a bit) can do it, but probably a very hard thing to accomplish. I'd most likely end up doing it with liquify in photoshop or creating controlnet models and praying hard xD
lets be honest, nearly no one looks close enough to figure out errors in pictures nowadays... same reason its not that important to look after too much Details. also.. no picture is perfect. may it be an real Human or an Machine .
Thank you fro this video Mr.Talent. I don't really need it, but you made me sub for how indepth you made this. I'm a 2D artist that wanted to give AI a shot since it was making me scared for the future of art. But I really enjoy using AI. And is not as easy as people claim it to be. There's also such a large helpful community with people such as yourself giving in-depth videos and articles on workflows and methods to achieve certain things. Gonna sub as i definitely enjoy watching your very informative videos! :)
Thank you so much for the kind comment! Really like seeing artists getting into AI and being a part of the community, hope you continue to find enjoyment on it :3
@@Not4Talent_AI I will continue to support your channel. It's incredible how much i've learned aside from complex hand posing. I even learned some inpainting tricks too. Gonna try to finish a image i was trying to create with what i learned.
I learned that the AI looks for a solid reference to make the hands correctly, and that little bone that we have before the hand, This tells the stable diffusion what orientation it should take. 😎
Without that bone he is like Adam on Mother's Day. 😆😅😅😅😅
When the AI does not do what it is told, put that order as close to the first as possible and then if it does not pay attention, indicate 1.4 so that it obeys. It is possible that your data store does not have the requested reference, perhaps it is in another name, such as: sphere.
great insight, thanks for sharing!! @@canaljoseg0172
I must hand it you. Hands down, this is the most handy hand overview I've seen.
Omg 😂😂😂 you compensated all the puns I didnt put in the video. I had a handful too hahhahah
I have been clicking around on tutorials and guides to improve my results with SD and yours are by far the best I've seen so far. Concise with no fat or filler, comprehensive showing different ways of tackling problems, and you do a great job of breaking down complex concepts into steps that even a complete beginner can understand. Keep up the great work!
Thank you so much, comments like this make my day and mean a lot!!! Happy that it helped!
Bro you are wild, High quality video and making own tools with GPT that work while being useful. Mad respect!
Thank you!! Hope they help, Ill see if I can update the tool with some suggestions people send 😂
I learned a small fraction of these techniques a few months ago and then promptly forgot them through lack of use. This video seems to be PERFECTION for creating the best hands right now. Hopefully the state of generating hands wil improve but THANK YOU so much for your efforts, not least your webcam tool.
hahaha I hope so too tbh, quite a bit of work atm xD Thank so much for the kind words, hope it helps!!
You really go above and beyond to create the most comprehensive video possible. Between your last video and this video, I switched to ComfyUI since A1111 just broke one day and wouldn't generate anything other than noise (or just give the nan error). I thought I wouldn't be able to follow your tutorials anymore, but there's so much info here that will even help me in ComfyUI.
Thanks for these super detailed videos and please keep up the good work.
Also wishing you a speedy recovery, take care.
Thank you omegasuperveryfking much for the kind words. With comments like this it's so much easier to create content. Hope it helps and, maybe we could try to see why A111 was crashing. Have you re-installed?
I'll try to keep providing value for as long as I can!
@@Not4Talent_AI
I did try reinstalling it but it didn't work. I also made a reddit post about it asking for help. (I initially put the link here but then TH-cam kept deleting my reply for the links so had to delete it)
After reinstalling, things became worse for some reason. As the one or two generations I was getting before were also replaced by nan errors.
I changed my nvidia drivers to studio drivers instead of game ready drivers to see if that helps but it didn't (yes I performed a clean installation).
To be honest, I think it's mostly because my laptop is potato (4GB VRAM) considering image generation standards, but I'm just confused as to why it worked before.
After switching to ComfyUI, I do miss all the extensions from A1111 but it works. There's just more I have to do to get the work done. Also, ComfyUI uses a lot of CPU RAM in lowvram mode so I think that's the reason it works on my laptop. From what I've understood, it offloads the model onto CPU RAM whenever they're not in use. It also uses CPU for latent generation (maybe? not sure. read it on reddit).
As a result I can generate a 512x512 image in about 30-40 seconds compared to A1111 which took about 2-3 minutes.I still wish I could use it though, cuz it makes some things way easier, so a combination of both would be my preferred way.
Sorry for the long reply 😅. Hope I didn't take too much of your time. Please just continue to do what you do. Your videos are more than helpful no matter what I use since a lot of what you tell are fundamentals of image generation. I look forward to your future content!
By far the best SD creator!! I check for your content daily, keep it up, nobody makes content on your level for people who are familiar with the tools SD has to offer.
Thanks a lot, really! Tbh this video is a lot of manual work, but I hope I'll find better ways as things improve. but thanks again for the support, means a lot
@@Not4Talent_AI for someone like me, who is not artistically inclined with their hands, your toolset and creativity has helped me a lot.
Great job visually showing how to use this technology more artistically.
thank you!!
you are best sd creator when it comes to complex task.
Thanks very f much, means a lot!
I just discovered your channel and I really like your tutorials so far, thanks and I subbed :)
But what's really refreshing is to see an artist embracing generative AI instead of foaming from the mouth about it. Really shows what a great tool Stable Diffusion can be for artists.
The rest of the tutorial channels I'm subbed to are mostly tech geeks who've been using SD since its release and have gotten a lot of experience this way, so yours was a nice change :D
Thank you very much! that is ultimately my goal. Tring to use AI as a tool for artists to use ^^
By far one of the top SD content creators that actually is worth. Can't wait you to start with SDXL. If you start, haha. Cheers!
Thank you very much!!! I will start now that Controlnet Is finally working on there, even though I've been avoiding it cuz my gpu doesnt like it very much xD
The best tutorials on this channel: to the point, innovative, and well-explained. I really missed them. I hope you get better!
thank you so so much!! (for both things hahaha)
WOW a new vid.... I missed your tutorials. Keep it up brother. much love from Morocco :D
Much love from spain!! Will do, thank you!
I encourage people to go into any illustration software, and just do back and forth editing between manual edits to guide the generation explicitely and prompting with controlnet. You can get amazing results with just these 2 simple steps and some training! The more you do it, the better your images will get and so will your artistic skills (and it's super rewarding when it finally does what you want! haha)
Today more than ever, I value my talent a lot. my ability to draw what I want however I decide 😆😆😆
Every line is exactly under my control.
Yesss, being able to draw is very very nice hahahaha. Im still learning but well get there 😂
Bro you're insane, thanks for such a great tutorial
Thanks you for the kind words!
Long time no video ? The wait was worth it ! Like always, your videos are ... amusing and ... informative !!! Bravo à toi 👌
Thank you so much! Yeah, I fked up my hand and leg so I was AFK for a wile. Now back at almost full strength
@@Not4Talent_AI Oh! that's why, well just don't screw them up next time and you will be okay 4Head xD.
hahahahhaha I'll try not to fall into a broken plate for as long as I can xd@@Dex_1M
I'm 2:43 in and honestly you still draw better hands than me even when they're like that.
hahhahah ty! I actually draw a bunch of hands from time to time. that was a quick sketch tho. But they are so fking hard to do right 😅
Excellent video. Fast, to the point, and with tons of useful information. Exactly what I've been looking for. Thank you very much!
glad you liked it!! Ty for the positive words!
happy to see you posted again dude, stay healthy 😁
thank you!!! I'll try my best xD
0:18 hey it's me lmao. Thanks for using my pose as an example
Oh, really? Hahaha ty for posting loras and helping the community!!
Bro you are awesome, really this far the best video AI explaining, thank you.
thank you so much for the kind comment! hope it is useful
Int's just another level! Muchisimas gracias, hombre!
jajajaj gracias a ti!!
These tutorials are amazing! Great work! Some of the most fine-tuned videos I have seen
thank you so much!!! hope they help!
Oh yes hands.. The Archenemy of Stable Diffusion. For me it also got better with SDXL but still it's my createst enemy.. Thanks for this helpful video 😊
True, I didnt mention it... sht... Sdxl is better with hands, so as controlnet gets better for it, it will be easier to make them there.
Even though atm I couldnt really get good generations using CN and SDXL so I left it out.
Still should've mentioned😂😂
Thank you!
@@Not4Talent_AI Well for my last model DEMON CORE I needed to create a picture with a woman holding a metal ball. It did it well but I needed about 10 trys to get it perfect without using inpainting. There were mostly small errors like holding the ball from or a strange position of the finger. Overall it did quite well ;-)
Your content is always packed with the BEST tips and tricks my dude. Thank you
super happy to read comments like yours! thank you so much, hope it helps :3
the goat is backkkkkkkkk, epic video, p.s. for my workflow I send images to the photopea extension instead of using inpaint sketch to be able to fully edit the canvas easier.
Thats a realy nice idea too hahahhaha Thanks btw!! we are backkkkk
Thanks soo much @Not4Talent, you got me giga interested in Stable Diffusion, and I'm seeing great progress soo far! you're not just a great teacher you're also entertaining to watch
Super happy to read this!! Thank you so much for the kind words! Hope you find it enjoyable and helpful :3
Thanks for the tutorial!
ty for the kind comment!
Really good tips, I like that you just use photoshop when it is easier, good images are just all about Stable Diffusion they are just all tools.
thanks!!! Hahahaha I agree, treating SD like a tool to help is 100% the philosophy here hahah
You are amazing bro, thanks for all the effort put on this!
Ty so much! Hope it helps
Amazing tutorial
thank you so much!
the AI still needs a lot of adjustment and trial and error to get good results. I think you did a good job fixing the mistakes by hand!
Unfortunately yeah hahaha ty btw!
Patience = Power 😉 this is why I *love* manual work. More accuracy, control and quality. Now that I look at it, it's not really *that much* more work, looks totally fine.
If I were to try prompt-only with no cnet and tweaks, I'd lose all my hair after the first session of a couple hundred images
But I get it, people just want to get professional looking jackpot image in minutes ¯\_(ツ)_/¯ this is why we will stay on top of the quality work
hahahhahha fast dopamine > good dopamine over time I guess xD
Awesome work ❤
thank youu!!!
Thanks for your dedication and effort, really appreciated!
Ty for watching! Hope it helps. (btw, lora all the way. Cn can help but it will not make consistent characters)
Awesome! Thank you!
Ty for watching 😄
Great tutorial! Thank you, keep at it. Peace from Cape Town
thank you!!! Peace from Spain :3
It's unbelievable, how much time did you spend researching that problem? Thank you very much.
tyty!
There are some faster ways to do this with just inpainting and a bit of luck now. Mainly with anime and Ponyxl. Or using SDXL with models that are nicely trained.
Might want to check that too if you struggle with some hands
research probably bout 1 week or so, practicing a bit more
amazing tutorial!
thanks!! hope it helps
The best advanced AI content creator in the game. Better than the clickbaity 100k ones that only show beginner's content.
hahahhahah thank you so much! hope it helps!
Is it just me who thinks it's really funny he said he "messed up his hand" IRL, while recording a hand fix tutorial :D
Pls dont mess up your hand ;)
hahahahahha I'll try not to mess it up again xD. Also fked up my leg in the process
Very cool techniques! I noticed that your videos pack in a ton of info at lightning speed. It would be awesome if you could do some 'part 2' versions of videos like for this one where you do one or maybe two examples of the technique from start to finish. This could be good for a live stream that gets saved for later viewing too. So you wouldn't even need to edit the video which I'm sure takes a ton time.
thank you!! I've been thinking of a way to do that but I still havent figured it out. I'd like to keep the video's on a steady standard when it comes to editing and style, so mixing it with long unedited format videos feels weird. Also, I would stream but my internet is bad af... So yeah, I really agree with what you say but I have no idea how to go about it xD
The suggestion is pretty nice tho, I have to think about it more. Tyty!!
Dude why are you so underrated?? I have watched tons of AI content creators and tutorials and no one explains it with so much passion and accuracy. Truly game changing. I have a question tho: What can I do if I need the same hand, but on other characters and in different positions? I guess the answer is "train a LORA" but is there any other way? If not, can I train a LORA just on hands? I also need to train the LORA on both hands.
thank you so much!!
I'm not sure I understand the question tho, do you want to use the same hand pose on multiple characters? Is it in the same position always or does it change position but not pose?
You can train hand LORAs, like this: civitai.com/models/99237/hand-on-own-chin-or-pose
(it doesnt give amazing results, but it's very fast and makes a base image that's easier to clean up than just having to star all over again for a new character)
If the position doesnt change, just the character, I think hands should be similar enough where you can just change skin tone and reuse hands.
If position and character change, I think it depends on what the goal is, but You could create a library of openpose + depth + edge references to re-use and pose properly with something like the "depth library extension". then just remember which weights work best for the checkpoint you are using
Hope this helps!
holy shit. that's an awesome tutorial!
thank you so much!!! Hope it is useful hahahah
I guess in the time that all the manual work takes, you could just have created 300 other images more or 300 variations of the one you like and check if one of these got the hands right. ;-) I've also the impression that the bad-hands5 embedding makes it at least a little better. I saw it at least trying to fix fingers multiple times. But let's say you finally got SD to get the hands right. Now you want to upscale the image and in that process it has to be created again, that is if you actually want new details and not an image that is just a lot bigger. Now you can choose between a blurry image that is forced by c-nets to be as close to the original as possible or a nice new detailed, similiar looking image that got the hands wrong again... after minutes of computing. Maybe it's just me, but there are days, where I want to have nothing to do with this stuff anymore. ;-)
If you have the computing power then yeah, maybe you can do that. But I think having the manual tools to fix stuff is straight up more consistent.
I agree that upscaling can be tough, even though not as big a problem as you make it look like. (imo).
And yeah, AI is not as easy as it looked in the beggning, at least not if you want actually good images that look professional. Which is why I'm a huge advocate for manual work, since it lets you learn and overcome this issues while AI advances
"for people with this many hands they aren't good at holding balls" he said it guys
Very good, good rays to you!
ty!!
9:50 Little pitfall here: If you use inpaint sketch or Inpaint for small things / small masks,
you need to reduce the "Mask blur", to not lose your change in the too large blurred area.
So to get those thin fingers inpainted, you would need to lower it quite a bit to actually get them generated, depending on the resolution of the image.
If it's 512x512, you would probably have to lower it to below 4 here to actually get a finger. In 1k x 1k you would maybe need 8 or even less.
I regularely miss this and get frustrated before it plops up in my mind again that my blur is too high.
true, forgot to say that tbh. most of the time I use like 1 or 2 blur for precise stuff. Ty ty!!
you are a genius
Hahahhahaha thanks!! (Not that I agree, but appriciate it 😂)
as an artist myself I tempted to use SD like this. but i think it's more like using SD as replacement for pinterest. openpose is good. I definitely will explore that. and ultimately train my own characters later.
Nice!! Hope that you enjoy it!
You are a fucking beast bro, keep it up this kind of videos, i'll start watching all your content related to this, thank you very much!
thank you so much!!! hope it helps :3
4:00 oh yeah there is Paint toll sai. it has some really good Linework for vector use. you can trace over it and even edit and add pressure to the lines. if you only have mouse. I feel like you know what I'm talking about.
Tbh I had no idea hhaahha thats some nice info right there xD tyty!
You're so good with Photoshop, I understand you need AI to create 60% of the image then, the AI will hold your beer while you do your magical tricks. Your workflow is super interesting for lamers like me. On my side, I just generate zillions of images with diff prompts technics, Lora's, checkpoints and TI (even if I know they make hands worse usually) then throw it to img2img and hope I'll get the jackpot without be like "inpainting" anything.... and of course, each image is saved on my side, just in case of, "you know", lol... throbbing the perfect ticket. I guess, it's a two school vision : yours, you'll have to put efforts learning something else, dominate your tools, dig a stone and fine tune it... mine, just blow it up and sure, you'll be surprized, at some point but you can't be sure you can reproduce it... and it was a lucky day and the lucky days are not common... shrug
Thank you for that compliment ahhahhah. I have been using it for a long time but, tbh, I think most people can get to a level where you can do this in about 1 or 2 weeks. So if you feel like it, I'd encourage you to try it if it is something you will need. If not, your approach works too xD Iterate until the right answer comes up ahhaahha. Ngl I do that too whenever I'm not focused on creating something and just want to generate stuff on the side.
hope that, either path you choose, you accomplish what you wanted! Thanks for the kind words too :3
Excellent hand holding kidney my friend.
dont know if this is positive or a negative comment, but thanks just in case hahaha (mb, I just got it XD)
That was my attempt at a joke. I love your channel and content its great.@@Not4Talent_AI
hahahahha I'm just too dumb xD Thanks!!! @@twistedbydsign99
Bro... welcome to the club @@Not4Talent_AI
AMAZING VIDEOS ! PLEASE MAKE A CLOTH LORA TRAINING VIDEO TOO
Thank you!! I havent trained any clothings myself, but its on my list of things. Might take a while tho, srry for that
@@Not4Talent_AI that's fine . I will wait 😊
I only work with Midjourney but anyway your videos are one of the best sources of inspiration to solve problems related to posing and composition I've ever found. Thanks, man.
thank you! Never expected having subs that mainly use midjourney, interesting hahha hope something helps!
Amazing tips and tricks for a good hand (everyone want good hands!!!11) like + subscribed
Thank you so much!! Hope they help :3
Can you share what exactly prompt, denoising strength and cfg scale you've used while inpainting sketch? I'm only beginning to learn SD so i dont have idea about control net, but your results in inpaint sketch were pretty cool so i want to achieve the same.
oh, I changed pc so I dont really have the images. But I usually use different ones depending on how accurate the sketch was. IF I used a good color and a nice shape, I'll probably keep it with a mask blur of 1 or 2 and a denoise of 0.2 to 0.35
If I did a bad sketch that was more of a "base" for SD to iterate from. Then I'll use a higher blur and higher denoise. Like 0.4-0-75
CFG I usually just move between 5 and 8. Mainly 7 .
Prompt you always prompt for what you want to see inside the mask, unless you also want to give some context about the rest of the image.
I wonder if you could make an extension that used a GAN (adversary network) to regenerate the inpainting of a hand until it looks like a source image, so you could just walk away and let the software to the generating until something good comes out....
That would be pretty cool. Hope someone with the skillset tries to do that tbh
which model do you suggest for photorealistic hand inpaitining? thanks a lot
I usually use epicrealism
Hi there. Great video, although the answers and methods are great. But that's only a single person.
It sadly could not help me with MY specific problem. I am trying to use CN will all its possibilities, mainly OpenPose, Depthmaps and Canny to generate images with dynamic poses with TWO (or more persons) closely interacting with each other. For example as "simple" as a couple dancing. But I can't get any results, even with your tips here on different CN models. I also used PoseMyArt from the beginning. But whatever I tried in the last couple of weeks, months even: Either, the result is butt-ugly or SD (or rather CN?) simply does not what it is asked for.
If you know how do get something like this done. PLEASE, PLEASE do a tutorial on that.
Hi!! thanks! that's literally the next video. If you have a specific image you want to see done, please mail me a sketch or something and I'll see if I can add it as an example for the vid. I'm currently scripting so I think I have most of the workflow layed out
Thanks but no thanks. I wouldn't have the patience... I think I will make them wear mittens instead. I run out of memory when I use two controlnets anyways. but thanks for sharing your experimentations.
hahahahha thats completely fair XD ty for the feedback :3
@@Not4Talent_AI hopefully a miracle solution will pop up and you'll tell us about it. Some whiz could come up with a script like LH(50,0,100,100,100) -> Left hand thumb bent 50%, index straight (0% bent), other 3 fingers bent 100%. Idk...
xDDDDDDDDDDDDDD I pray for the day that comes hahahaha@@ravioli34
The problem with this is that this is actually working for your picture and then we wouldn't be fulfilling all of the stereotypes that no-ai protestors are spouting.
(Seriously though - thanks for the compilation of techniques. I knew most of these already, but I somehow never came across the idea of just painting my own canny controlnet...)
hahahahahha true xD
Thank you! Glad it was somewhat helpful :3
2:28 With a little bit of practice, you won't need AI in the future :)
hahahhah thank you so much!! I still practice just for the fun of it tbh xD
@@Not4Talent_AI You should try more ^^'
ty! I will :3 @@losing_interest_in_everything
After drawing the hand in photoshop (11:03), which pre-processor should I use? I found it a bit confusing.
Should the background always be black?
I tried to do it but it was messy, the edges of the hand went into the image, it was strange.
I'm using the "depth map library" to position the hand on the image, so I don't know if there's a problem.
hi! none should be the preprocessor. So literaly no preprocessor should be used if the lineart is done properly.
I guess using the hand library extension is fine to use btw. If with "none" it still fails. Maybe try making sure that it is properly position with another editor like photopea or photoshop
@@Not4Talent_AI I finally succeeded, but I had to use a high denoise strength of 0.75. Last question, is it possible to draw "perfect" nails with this? Thank you.
I'd guess if you draw them properly yeah. You can always get help with inpaint sketch @@rAtableable No problem!
@@Not4Talent_AI Thanks.
Very informative as always, but what I see it's still very very very time consuming to get good hand results...
Unfortunately, yeah 😂😂 ty btw!
OMG I needed it the most
hope it helps!!
It doesn't get any more complicated than that
hahahhahahhahaha, sure hope it doesnt xd
@@Not4Talent_AI Adetailer helps me well
Ngl, it never does shit for me xD No matter how much I play with the parameters it still makes trash hahaha@@SanftmutLP
There are some pretty big mistakes even in the thumbnail, but with some more perspective training and study from life you'll be perfect at hands! Keep it up ❤
yeah, tbh it depends on the person's skill. since mine aint that great, I cant fix every mistake. Hopefully eventually, thank you!!
FiNgErTiPsss !... FINGeeeerrrTIIIIPSsssss !!! 😲
FIIIIIIIIINGERTIPPPPPPPPPPPPS
Title: "How to do hands in SD"
Video: "We will fix it in photoshop"
😂😂😂 I mean fixing is in photoshop. Hands are in SD. But yeah there is a lot of manual work.
better cn models and optios have been released recently though. Now there is not that much of a need for fixing. Even if still a bit
hmm any idea for hand skin contact other character like squeeze or damage impact
havent tried tbh, sounds tough. Maybe prompting with inpaint sketch (trying to help aI a bit) can do it, but probably a very hard thing to accomplish. I'd most likely end up doing it with liquify in photoshop or creating controlnet models and praying hard xD
bro how are you so smart
hhahahahahhahaha I dont know about smart, more like stubborn xD but thanks :3
como te va con el Slider training video? Are you still working on it?
Yup! Im creating a webui for it so yall can train a bit more comfortably. So its taking a wile.
But hopefully ill have it by early next week.
@@Not4Talent_AI thats so cool. thanks for that. really looking forward to
bua, ns xk he respondido en ingles XD
Pero la tech mola un huevo, veré si puedo hacer un video a la altura@@___x__x_r___xa__x_____f______
@@Not4Talent_AI hahahaha
q mono el pinguinito ❤
supa
14:50 💀😂
You talk about "pose my art", but what is it?
app.posemy.art
Its a website where you can pose characters (and others) To later export maps like open pose, depth, normal... etc
😂😂 Says, "I can't draw," then proceeds to put up awesome hand-drawn sketches. 😂😂
hahhahahhaha tyty
I guess I meant that not good enough for them to be perfect, and less so with color and shape
Sakura, de naruto.
Lux, de league of legends.
nahhh no me suena
How to make the hand look natural
This video: Learn to draw hand
Pretty much 😂
kinda ironic that ur basically drawing here lol...
hahahahaah yeah pretty much. and drawing hands at that xd
at the end of the day that will always be the best way
and they said ai was easy 😩
hahaahha. right?
''I can't draw'' is essentially the reasoning behind many people using AI and then passing it off as original art lol... cheaters.
lets be honest, nearly no one looks close enough to figure out errors in pictures nowadays... same reason its not that important to look after too much Details. also.. no picture is perfect. may it be an real Human or an Machine .
Hahahhaha could be yeah
Lol how to fix your hands, photoshop.😂
In realistic, photorealistic models, thats not so easy
yup, that's way harder for realistic models. That's why you inpaint with a reference mostly. instead of drawing over it.
but it's still way harder
Thank you fro this video Mr.Talent. I don't really need it, but you made me sub for how indepth you made this. I'm a 2D artist that wanted to give AI a shot since it was making me scared for the future of art. But I really enjoy using AI. And is not as easy as people claim it to be. There's also such a large helpful community with people such as yourself giving in-depth videos and articles on workflows and methods to achieve certain things. Gonna sub as i definitely enjoy watching your very informative videos! :)
Thank you so much for the kind comment! Really like seeing artists getting into AI and being a part of the community, hope you continue to find enjoyment on it :3
@@Not4Talent_AI I will continue to support your channel. It's incredible how much i've learned aside from complex hand posing. I even learned some inpainting tricks too. Gonna try to finish a image i was trying to create with what i learned.
Thanks!!! Hope you get nice results! @@anaversary-
Weird, but interesting. 🤔
hahhahahha ty! I guess 🤔