These tutorial series has really SHOWN (and not just said) how powerful Comfy is. People know it is powerful but never really got a deep explanation like these series. Mateo, true legend. Don't need to watch other youtubers regurgitating stuff other youtubers have said. THESE are the true tutorials!
I'm coming from A1111/Forge and trying like hell to learn Comfy. Your videos showed me just how insanely capable comfyui is and helped me understand what's actually going on in the process of generating an image. Can't believe I delayed learning this UI for so long -- it was daunting, but I finally feel like I'm beginning to understand it. Very valuable content!
Going to add one more comment here about how much I learned from this! The CFG rescale trick is going to be really useful, I had no idea it could improve multiple conditioning like that. Thank you ^_^
Very well a PRO, and your slow voice is a PLUS ! Then when you add some nodes, you do that a bit too fast, even we can go back to watch clearly. But a very good full of tricks Video ! TY
Amazing, most simple character tutorial ever. please make more tutorial about multi-IDs . Like : - making 2+ characters fight scene - putting the characters in a portrait with "multi area conditioning" nodes - applying instant ID to objects that can be held by characters (swords, smartphones, etc.) Thank you so much man, your instant ID just solve some character consistency problem in comic creation community.
im an a1111 user. these connecting nodes, adding bunch of different control-net models just scaring me. I see that is another level. However thats a perfect tutorial! Congrats my friend.
Thank you again Matt3o for another Best-in-Class tutorial, as a developer and as a ComfyUI teacher! You have no equal in this space. Even if there are also a few other good channels, your's IS THE BEST. Period.
take care of yourself, because, as I'm sure you can tell, there's an incredibly positive response to your content. it's probably exciting for you but please take it easy and don't burn out because you're just THAT much better than every other content creator on this topic. we need you! lol
instant id works with 1.5/v2 models, if you DON't use the loras, make sure the model is LCM compatible and use LCM lora, cache/retrieve your embeddings with v2 turned on (even if its not) and use LCM sampler, ddim_uniform/kerras scedular (i'm using it right now, and have been for a week) ;)
Thank you for your help. It would be slightly better if the video had more zoom (lower resolution), enough to identify the model used on a cellphone screen. Thank you, I couldn't find the solution until I came across this video.
It’s really difficult not to learn something tremendously useful by watching your master tutorials. It’s simply amazing how much knowledge you own from IA world and especially from ComfyUI. I have no words to thank you enough!
Let's generate a lot to see if we can have better hands ... (no-hands-picture-shows-up) Oh this is nice ! ;p Hands are so uncool in AI generation. Love your videos, it's just terribly hard to understand when you start with AI and comfyui. So in a nutshell, InstantID is cooler for its style purpose of the final image ? At the end I think that you have to learn every new thing that comes up and make your own choice. I'm still on SD1.5 as most of Fanart models are on that one in civitai, SDXL is not really my cup of tea. Cascade is very fast and nice but I miss my SD1.5 ... Waiting for SD3 and hopefully will be easier to make your own lora faster and easier.
@latentvision I’m not sure the batch images node actually would make a difference in the sampler, because the sampler input is an image, not an image batch. So in the end, I think all this does is cycle through the batch, but in the end, it would sample whatever the last image was in the batch. I’m guessing in order to truly end up with a little LoRA, one would have to figure out a conditioning weighting batch operation. That would be tremendously useful, but likely result in heavy resource load.
At 5:18, you say that there are far better models for photorealism. Can you recommend any for Comfyui in SDXL? I have a single photo of someone and can't create a Lora from that, but want to use it for generating other realistic photos.
Great work. Please guide me to install DLib to be used for ComfyUI instantid. 1.Where to create the folder in the ComfyUI folder 2. How to install then
It's a very amazing tutorial, thanks.❤ I wonder why in 11:50, you used "condition combined" instead of chaining "positive and negative" conditions of 1st InstantID to 2nd InstantID? 😅
instantid is a kind of uncommon controlnet, at the moment that is the best solution, I need to check if I can send the text prompts directly into attention, that would simplify things
Another really interesting video. If you don't mind me asking, is this all in python? I'm learning how to code and it would be interesting to try to apply some of it to comfyUI to earn more about AI on my coding journey. Love your content, keep it coming!
Matteo, could you cover such functionalities like Patch Attention and the Controlnet Attention? I think only you can help us with it and show the use cases for that, would be very interesting.
It's me again, with a question this time! You are using an image for masking (at 4:48) and you keep the same image through the whole video, even using an inversion mask for the 2nd IpAdapter. I'm not quite how this works : are you supposing the mask you are drawing will more or less be where the face of the mona lisa will be, to take only the rest of the painting? Wouldn't it be simpler to c/c the reference image, draw a perfect mask on the Mona Lisa''s face and invert it? Thank you!
After installing the package, I cannot access the nodes needed for the program to work. Being human and new at this, I feel that I may have neglected a detail or managed to put something in the wrong place. any assistance would be greatly appreciated.
very good tutorial, but the custom nodes mentioned installation are very hard and painful to not have errors, thus you can't follow the result on this tutorial. im stuck on faceanalysis
You are an absolute master. I've learned so much. However I still can't understand why all my output is a face close up despite trying to show "three-quarters view" or full head and chest or even full body. What am I doing wrong? Thank you!
You mentioned InstantID is to add styling to someone's image. And you said there are far better photorealistic face swapping models. Which are the best ones in your opinion, to do the best, most realistic faceswapping? Thank you!
Amazing video. How to do when we are not able to pass under the 1.2 for euclidean and 0.8 for cosine? I used your workflow. It works perfectly with your inpu images but with my personal inputs, I don't have the same results. My 3 images for the InstanID and the IPAdapter are wide and the face enter in a 640x640px...THANK YOU in advance.
Thank you for the most useful information on comfyUI. Your lessons are out of competition. Unfortunately, I still can't figure out why there is some kind of noise at the output (without any errors in console), noticed that if you connect the clip text encode directly to ksampler bypassing instantid, then everything is fine with the image, but the face is not too similar. And I noticed that it works much faster than other instantid before him, even faster than Forge in webui
@vision Thx!! Update really solved the problem. It works very fast on my 8gb graphics card! Now it really works as regular generation instead of 10+ minutes
previus error was solve downloading all models with instantID... but new error came when try the ksampler run Error occurred when executing KSampler: 'NoneType' object has no attribute 'shape' SOLVED: need an XL checkpoint... srry.
I have to say InstantID is quite effective, especially for creating side faces from only frontal references, which the other face models often refuse to draw. I do run into a weird problem though. It appears my InstantID workflow wants to work at much lower CFG (~2) than what the model would usually prefer (4~6), so I need to either sacrifice background complexity by running very low CFG or risking burning the image with high CFG. I wonder if this is specifically an InstantID issue?
@@latentvision I guess the problem is that the checkpoint normally works quite well at CFG=5, but when connected with the Instant ID node, it produces images with weird colours unless I turn down the CFG...
@@mithrillis CFG 5 is very low already. InstantID requires CFG 4-6 if the checkpoint works at 7-8. So if the starting point is 5 it makes sense that you need to lower it to 2-3. Again try with rescalecfg
Mateo I'm thinking this works best for advanced users of comfy. I'm still reviewing the set up, which is quite complex. However you get what you put in when it comes to the analytics of the spaghetti pipelines and TBH you really are in control of what you want as the final out put. Not for those looking for a quick fix in the production of AI rendering. I remained subbed (for the learning experience and nuance)
@latentvision Dear Matteo , I am trying to use Depth-anything (Marigold depth estimation) within your InstantID workflow like the one used in the InstantID Huggingface space , but unfortunately I am facing error because the depth-anything model is based on SD1.5 , could you please help me about that , I can give you my workflow to check it . And as usual your tutorials are very helpful and easy to learn from it , Thanks a lot. 👍
I wonder how do I upscale the art without losing the face likeness from 1024 to 2k or even 4k? I tried a couple of methods but the likeness are totally gone even at low denoising str.
Amazing!! Will there be an implementation for sd1.5? These models for SDXL are flying way over my 8gb vram system. Tried it on colab though. Vram tops at 14 gigs.😂
hello got a problem with my amd 7900xt comfy ui gives me an error. Error occurred when executing KSampler: Could not allocate tensor with 20971520 bytes. There is not enough GPU video memory available
What kind of GPU or CPU you're using? Cuz for a simple example I'm already waiting 18min and it's on 87% :D it will take forever to play around with prompts.
I don't have the "Apply IPAdapter" or the "Load IPAdapter model" nodes and I have IPAdapter install and I see lots of other related nodes. Had something changed in recent versions? am I missing something?
@@latentvision Thanks! I thought I was losing my mind, I spent half of the day trying to figure out why I don't have the nodes everyone is using. Looking forward for the documentation.
Hi! THX for you awesome video! Quick question. You used "conditioning combine" to link multiple InstantID in parallel, but what about connecting them in series instead?
each instantid controlnet instance has to be patched with the person's embeds. If you connect them in series the embeds get averaged. I'll see if there's some sorcery I can do to make it work, but at the moment this seems to be working.
Hi Mateo, in the last Apply instantID advanced part, you said that we can choose to fine-tune ip_weight and cn_strength. Do we set the continuation degree of the reference image style by adjusting the ip_weight, and set the proximity to the facial features and prompt words of the reference image by adjusting cn_weight? Thanks.
These tutorial series has really SHOWN (and not just said) how powerful Comfy is. People know it is powerful but never really got a deep explanation like these series.
Mateo, true legend. Don't need to watch other youtubers regurgitating stuff other youtubers have said. THESE are the true tutorials!
Love it. Learned more in 15 minutes than I have in a month of struggling with basics. Thank you!
God, there was so much to unpack in this video...I'm going to be coming back to this for months to understand it.
I'm coming from A1111/Forge and trying like hell to learn Comfy. Your videos showed me just how insanely capable comfyui is and helped me understand what's actually going on in the process of generating an image. Can't believe I delayed learning this UI for so long -- it was daunting, but I finally feel like I'm beginning to understand it. Very valuable content!
We respect and support every lecture you give.Kudos to you for your overwhelming knowledge.
Going to add one more comment here about how much I learned from this! The CFG rescale trick is going to be really useful, I had no idea it could improve multiple conditioning like that. Thank you ^_^
Very well a PRO, and your slow voice is a PLUS !
Then when you add some nodes, you do that a bit too fast, even we can go back to watch clearly. But a very good full of tricks Video ! TY
Don't have the time or motivation to dive into this, but a tutorial of this quality deserves kudos even from someone like me. :)
By far the best ComfyUI tutorials on YT. Thank you so much 🙏
100%
100%
This guy is the goat- writes the nodes and gives useful and detailed instruction on how to use them.
Agreed
I also agree.
100％
The man knows how stable diffusion's brain works, your the best instructor I've ever seen Mateo, keep the great work
Thank you Mateo, your channel is definitely my favorite UA-cam channel on the topic.
Amazing, most simple character tutorial ever.
please make more tutorial about multi-IDs . Like :
- making 2+ characters fight scene
- putting the characters in a portrait with "multi area conditioning" nodes
- applying instant ID to objects that can be held by characters (swords, smartphones, etc.)
Thank you so much man, your instant ID just solve some character consistency problem in comic creation community.
im an a1111 user. these connecting nodes, adding bunch of different control-net models just scaring me. I see that is another level. However thats a perfect tutorial! Congrats my friend.
Thank you! I finally got this working as I followed along. Lookin forward to more :)
Thank you again Matt3o for another Best-in-Class tutorial, as a developer and as a ComfyUI teacher!
You have no equal in this space. Even if there are also a few other good channels, your's IS THE BEST. Period.
take care of yourself, because, as I'm sure you can tell, there's an incredibly positive response to your content. it's probably exciting for you but please take it easy and don't burn out because you're just THAT much better than every other content creator on this topic. we need you! lol
LOL... okay, I'll take a break 😄
@@latentvision noooooooo too long ;) you are the GOAT!
Ты просто боженька чувак! Это невероятно круто! Твое видео в 15 минут надо разбирать неделю!))) Спасибо большое за твою работу!
Well, yes. You rock. You give us the *what*, and the most important *why*.
I spent an hour going through your video. thank you, very helpful.
I haven’t watched any comfy ui vids yet but dang this looks fun from an experimenting perspective
Whew! Amazing as usual. Thank you!
Another excellent video, thank you
Complimenti Matteo, every Comyui node that you do, is amazing :)
Thank you for your tools development and videos you make, it is amazingly helpful to improve skills and keep moving industry forward!
Amazing videos man. Love u.
I learn more from your videos than all other AI videos combined.
Your lectures are always fantastic!
Thank you so much for your hard work. Your tools and tutorials are incredible. You are truly awesome.
you are a genius teacher
Holy moly I learned a ton with your video. Thank you!
instant id works with 1.5/v2 models, if you DON't use the loras, make sure the model is LCM compatible and use LCM lora, cache/retrieve your embeddings with v2 turned on (even if its not) and use LCM sampler, ddim_uniform/kerras scedular (i'm using it right now, and have been for a week) ;)
what? how? Do you have a comfy workflow you can share?
what is v2?
@mordokai597 can u plz share it
Another great video packed with valuable insights 🎉 - Thank you Matteo!
Great work as usual, thanks!
best comfy related all time guide.
That's what I was waiting for! Grazie mille
Amazing work!
Thank you!
Y'all are correct. Protect this man at all costs.
LOL, thanks for the laugh :D
Thank you for your help. It would be slightly better if the video had more zoom (lower resolution), enough to identify the model used on a cellphone screen. Thank you, I couldn't find the solution until I came across this video.
Sei il migliore, grazie! Ciao!
Keep up the good work!
You are the best!! ❤
this is just brilliant. thanks a lot.
One day, I wish to be as incredible as you. Thank You.
Thank you very much 😀 that was helpful.
It’s really difficult not to learn something tremendously useful by watching your master tutorials. It’s simply amazing how much knowledge you own from IA world and especially from ComfyUI. I have no words to thank you enough!
Look like it's not working for me, with the latest version of ComfyUI (Portable 0.3.10)
Let's generate a lot to see if we can have better hands ... (no-hands-picture-shows-up) Oh this is nice ! ;p
Hands are so uncool in AI generation.
Love your videos, it's just terribly hard to understand when you start with AI and comfyui.
So in a nutshell, InstantID is cooler for its style purpose of the final image ? At the end I think that you have to learn every new thing that comes up and make your own choice.
I'm still on SD1.5 as most of Fanart models are on that one in civitai, SDXL is not really my cup of tea. Cascade is very fast and nice but I miss my SD1.5 ... Waiting for SD3 and hopefully will be easier to make your own lora faster and easier.
@latentvision I’m not sure the batch images node actually would make a difference in the sampler, because the sampler input is an image, not an image batch. So in the end, I think all this does is cycle through the batch, but in the end, it would sample whatever the last image was in the batch.
I’m guessing in order to truly end up with a little LoRA, one would have to figure out a conditioning weighting batch operation. That would be tremendously useful, but likely result in heavy resource load.
the face embeds are averaged
This is masterpiece mate 🙏
Thank you Matteo! love your work.
Very great tutorial. Thank you very much!
very informative.thanks.
Thanks !
6:22 : recale CFG
At 5:18, you say that there are far better models for photorealism.
Can you recommend any for Comfyui in SDXL?
I have a single photo of someone and can't create a Lora from that, but want to use it for generating other realistic photos.
Thanks so much for this ❤
this is fire 🔥 thanks!
Thank you matteo 😊
Great work. Please guide me to install DLib to be used for ComfyUI instantid.
1.Where to create the folder in the ComfyUI folder
2. How to install then
At 5:08 you said there are far better models for photo realism. What were you talking about specifically and what are the models?
check my "face science" video ua-cam.com/video/oBKcjY-JO3Y/v-deo.htmlsi=hF0zDCAuFlZn05sp
Thank you for sharing part of your knowledge :)
Can you make one for photorealistic results, please? with all the optimal settings
Also I have all the IPAdapter stuff installed but there is no Apply IPAdapter?
It's a very amazing tutorial, thanks.❤
I wonder why in 11:50, you used "condition combined" instead of chaining "positive and negative" conditions of 1st InstantID to 2nd InstantID? 😅
instantid is a kind of uncommon controlnet, at the moment that is the best solution, I need to check if I can send the text prompts directly into attention, that would simplify things
👍 That would be very effective. 🎉
I wait for your videos. I have subscribed to quite a few channels, but yours is the only one with 🔔 on. Thanks! 🙏
Great! I was looking for this on your channel, an hour ago!
You are a Genius!
Great video. I wish I could subscribe twice.
Another really interesting video. If you don't mind me asking, is this all in python? I'm learning how to code and it would be interesting to try to apply some of it to comfyUI to earn more about AI on my coding journey.
Love your content, keep it coming!
yes, mostly python (and thanks! 😄)
Matteo, could you cover such functionalities like Patch Attention and the Controlnet Attention? I think only you can help us with it and show the use cases for that, would be very interesting.
Amazing tutorial!! ...I have a question: how to make preview image for only the specify node (with out running Queue Prompt)?
I am getting a some loss of identity when using image kps for head positioning
do you have solutions for that? I have the same problem
It's me again, with a question this time!
You are using an image for masking (at 4:48) and you keep the same image through the whole video, even using an inversion mask for the 2nd IpAdapter.
I'm not quite how this works : are you supposing the mask you are drawing will more or less be where the face of the mona lisa will be, to take only the rest of the painting?
Wouldn't it be simpler to c/c the reference image, draw a perfect mask on the Mona Lisa''s face and invert it?
Thank you!
the face position is pretty much fixed and the attention mask is rather loose anyway, so yeah it works :)
Mind. Blown
After installing the package, I cannot access the nodes needed for the program to work. Being human and new at this, I feel that I may have neglected a detail or managed to put something in the wrong place. any assistance would be greatly appreciated.
very good tutorial, but the custom nodes mentioned installation are very hard and painful to not have errors, thus you can't follow the result on this tutorial. im stuck on faceanalysis
You are an absolute master. I've learned so much. However I still can't understand why all my output is a face close up despite trying to show "three-quarters view" or full head and chest or even full body. What am I doing wrong? Thank you!
with instantID you can set a reference for the pose
You mentioned InstantID is to add styling to someone's image. And you said there are far better photorealistic face swapping models. Which are the best ones in your opinion, to do the best, most realistic faceswapping? Thank you!
if you don't need a lot of control over the style (and facial expression) check my FACE SCIENCE video!
makes my beanbag tingle
I'm having issues with InstantID Face Analysis node, Could you please give me solution for this error? Thankyou
Amazing video. How to do when we are not able to pass under the 1.2 for euclidean and 0.8 for cosine? I used your workflow. It works perfectly with your inpu images but with my personal inputs, I don't have the same results. My 3 images for the InstanID and the IPAdapter are wide and the face enter in a 640x640px...THANK YOU in advance.
that's very hard to say without checking the actual images and workflow. Join my discord I can try to help there
Thank you for the most useful information on comfyUI. Your lessons are out of competition.
Unfortunately, I still can't figure out why there is some kind of noise at the output (without any errors in console), noticed that if you connect the clip text encode directly to ksampler bypassing instantid, then everything is fine with the image, but the face is not too similar.
And I noticed that it works much faster than other instantid before him, even faster than Forge in webui
you probably just need to update comfyui
@vision Thx!! Update really solved the problem. It works very fast on my 8gb graphics card!
Now it really works as regular generation instead of 10+ minutes
previus error was solve downloading all models with instantID... but new error came when try the ksampler run
Error occurred when executing KSampler:
'NoneType' object has no attribute 'shape'
SOLVED: need an XL checkpoint... srry.
I have to say InstantID is quite effective, especially for creating side faces from only frontal references, which the other face models often refuse to draw. I do run into a weird problem though. It appears my InstantID workflow wants to work at much lower CFG (~2) than what the model would usually prefer (4~6), so I need to either sacrifice background complexity by running very low CFG or risking burning the image with high CFG. I wonder if this is specifically an InstantID issue?
it depends on the checkpoint, if you need 2 it's seriously overtrained. Anyway you can use rescaleCFG
@@latentvision I guess the problem is that the checkpoint normally works quite well at CFG=5, but when connected with the Instant ID node, it produces images with weird colours unless I turn down the CFG...
@@mithrillis CFG 5 is very low already. InstantID requires CFG 4-6 if the checkpoint works at 7-8. So if the starting point is 5 it makes sense that you need to lower it to 2-3. Again try with rescalecfg
Which model is needed in the Load InstantID model node?
Mateo I'm thinking this works best for advanced users of comfy. I'm still reviewing the set up, which is quite complex. However you get what you put in when it comes to the analytics of the spaghetti pipelines and TBH you really are in control of what you want as the final out put. Not for those looking for a quick fix in the production of AI rendering. I remained subbed (for the learning experience and nuance)
check the "basics" tutorials if this is too much :)
@latentvision Dear Matteo , I am trying to use Depth-anything (Marigold depth estimation) within your InstantID workflow like the one used in the InstantID Huggingface space , but unfortunately I am facing error because the depth-anything model is based on SD1.5 , could you please help me about that , I can give you my workflow to check it .
And as usual your tutorials are very helpful and easy to learn from it , Thanks a lot. 👍
I wonder how do I upscale the art without losing the face likeness from 1024 to 2k or even 4k? I tried a couple of methods but the likeness are totally gone even at low denoising str.
hey there, i cant instal the custom node currently, any idea why? from comfy UI manager
Any recommendation for photo realistic model, I tried the same workflow with epicgasm timeline 3:33 but ksampler was failing.
realistic vision is generally good
Amazing!! Will there be an implementation for sd1.5? These models for SDXL are flying way over my 8gb vram system. Tried it on colab though. Vram tops at 14 gigs.😂
hello got a problem with my amd 7900xt comfy ui gives me an error. Error occurred when executing KSampler:
Could not allocate tensor with 20971520 bytes. There is not enough GPU video memory available
You're mentioning that there are "far better models for photorealistic results". Which ones are you referring to?
check my "face science" video!
will do thank you!! @@latentvision
👌👌
What kind of GPU or CPU you're using? Cuz for a simple example I'm already waiting 18min and it's on 87% :D it will take forever to play around with prompts.
can you use a mask and use Instant ID to just inpaint the face/hair area? so you can keep the background and person's clothes exactly the same
of course!
I don't have the "Apply IPAdapter" or the "Load IPAdapter model" nodes and I have IPAdapter install and I see lots of other related nodes.
Had something changed in recent versions? am I missing something?
yeah a new version was released today. I'll post the documentation later.
@@latentvision Thanks!
I thought I was losing my mind, I spent half of the day trying to figure out why I don't have the nodes everyone is using.
Looking forward for the documentation.
@@latentvision Hi! can you please tell me when you can publish the documentation? or have you already done it?
@@AntonRybalkin check my latest video :)
Hi! THX for you awesome video!
Quick question. You used "conditioning combine" to link multiple InstantID in parallel, but what about connecting them in series instead?
each instantid controlnet instance has to be patched with the person's embeds. If you connect them in series the embeds get averaged. I'll see if there's some sorcery I can do to make it work, but at the moment this seems to be working.
Hi Mateo, in the last Apply instantID advanced part, you said that we can choose to fine-tune ip_weight and cn_strength. Do we set the continuation degree of the reference image style by adjusting the ip_weight, and set the proximity to the facial features and prompt words of the reference image by adjusting cn_weight? Thanks.
they work together. ipadapter is a helper for instantid. instantid takes care of both the pose and about 70% of the likeliness.
Thanks for your quick response, I will think about it further more to understand it better.@@latentvision
I love your vids, But i cannot get indightface to work on my computer, any suggestions?
insightface is a tough cookie. check this thread github.com/cubiq/ComfyUI_IPAdapter_plus/issues/162
Cannot execute because a node is missing the class_type property.: Node ID '#11' How to fix it, thank you.