Image stability and repeatability (ComfyUI + IPAdapter)

[@tonikunec]

I feel like I am in some science class at the university watching your videos, you have so much knowledge, it's crazy. I watched all your videos and at the end of every one of them my jaw dropped to the floor… Amazing stuff! Keep up the good work, and thanks a lot!

[@smitty7326]

I'm not sure what it is, but these are the most clear and helpful guides I've found. I think it helps that you're both building it live and also explaining very clearly what each connection does. Thanks

[@goodie2shoes]

So many ComfyUI video's show very little and only hint at the possiblities. Your video's are a revelation! You show SO MUCH and have a deep knowledge of the interface and its possibilities. GREAT STUFF!! thankyou!

[@terrorcuda1832]

That was probably the best video on comfyui that I have seen. I'm new to comfyui and am facing the steep learning curve that goes with it. I find many videos overly technical to start with. You clearly explained your steps, not only how to do them, but why. I learned more from this video than probably the last 15 that I've watched combined! Thank you very much for your time. Keep up the fantastic work!

[@AgentofHappiness365]

I have to say... probably the best AI tutorials out there. Well done and thanks for the knowledge!

[@vladimirtsarapkin2840]

It is so uncommon to listen to a guide were author has such a clear understanding of the subject. Find you work extremely inspiring and helpful. Thank you, Matteo!

[@daniel99497]

U should win some kind of award for being so helpful to the community

[@omarei]

Yeah seriously

[@stephantual]

The irony is that it's him who's giving the awards. Matteo is an absolute legend 👑!

[@sifeyanis591]

Yes indeed ❤

[@ttul]

Really excellent overview of how you can put many node types together to achieve a fantastic result!

[@sanbaldo]

grazie matteo, questi video sono GOLD!! grazie per il tempo che stai dedicando alla comunità di comfyui!! super apprezzati (se mai ti venisse voglia sarebbe bello avere un corso 101)

[@JudithAlvarez-h6n]

Thank You so much, helped a lot. When I was teaching people, I never put music in my classes - people might want to concentrate on your words without music or listen to the music they like.

[@chadhamlet]

Wow! This is exactly what I was looking for. Thank you so much for creating this walkthrough!

[@Matt-st1tt]

Just thank you. Omg been looking for tutorials on this where I can open a work flow and look it through and follow along. Thank you.

[@jiucki]

Wow, thank you very much for this amazing tutorial! I'm definitely going to try comfy UI as I would like to have characters to create a small comic

[@McxCZ]

Jesus holy moly. Its like "So this 50 nodes is just a beggining".

[@amkire65]

Fantastic video. I have no idea how long I'd have been stumbling along on my own to try and do something similar.

[@sylwestermoniuszko-szymans1488]

few months back I thought the image generation AI is just type some text and get the results, which was boring, but fortunately people found a way to make it skillfull again

[@MrPaPaYa86]

Upvote if you are against background music. Great videos my man,keep up the good job this is one of the best sources on the topic

[@latentvision]

no worries, BGM has been removed long ago :)

[@swannschilling474]

Your videos are insanely dense and totally on top of being comfy to watch and packed with quality info!! ❤

[@Graino69]

Your tutorials have helped more than anything else on the internet, keep it up man! P.S. Ty for IPAdapter :)

[@pn4960]

Ciao Matteo! Thank you for your work, this is a very nice video. Easy to follow

[@othoapproto9603]

Food for thought, I couldn't eat another byte or bit. Thank you for sharing your skills and hard work.

[@WuduDionysia]

this was exactly what i have been looking for the whole time!!! thank you so much

[@WhySoBroke]

Supreme amazing stuff (like always) maestro latente!! No other person offers such great knowledge!!

[@jiml5166]

Wow. Using your tutorials I can pretty much manipulate any image to be what I imagine in my head. Thanks so much.

[@JuanS_DuodecimStudio]

Thanks Latent Vision, really helpfull, i' was a bit scared by this V2 changes, but it seems that is a good upgrade

[@Clupea101]

Excellent Guide, Top Man

This video is an international treasure. Protect it at all costs!

[@flisbonwlove]

Great work matt3o 👏👏👏

[@rockapedra1130]

You are a virtuoso comfyUI magician. I am in awe.

[@Milo_Estobar]

Very good, i knew from my very gut, that comfyui is for advanced user.... But never be able to see what its strength because most tutorial example keep giving easy example which makes me think "why do you even need comfyui?"... Now I have seen the full picture, thank you very much....

[@ashokp9260]

Awesome! Ur ideas know no bounds.

[@lucagenovese7207]

Come sempre, excellent.

[@TheRMartz12]

you are literally a GENIOUS

[@ImmacHn]

Oh God, I'm glad I came back to this video, I started using attention masks but misunderstood what they were for, I thought it was so the ipadapter would only take into account whatever was in the mask /_ \

[@lenny_Videos]

Awesome stuff 😊

[@pk.9436]

perfect work, thank you for inspire us 💪

[@KooroshGhotb]

The way you harness the power of ComfyUI is amazing, I wish you could make a tutorial for Comfyui, not just IP adapter. I'm sure there are a lot of things that we can learn from you. Please consider this.

[@latentvision]

I've made a couple of more generic tutorials, I'll do more in the future. thanks!

[@KooroshGhotb]

@@latentvision No sir, thank you for sharing your knowledge with us. Hope I could learn more from you.

[@KooroshGhotb]

A quick question though, I followed the link you provided in the captions for ComfyUI essentials extension, seems it leads to IP-Adapter GitHub page. Yet I couldn't find "Image Crop+", even after loading the workflow, Install missing custom nodes couldn't find it.

[@latentvision]

my bad, wrong link, should be fixed now@@KooroshGhotb

[@KooroshGhotb]

@@latentvision Thanks

[@GForcenuwan]

Amazing! Thank you Matteo

[@KriGeta]

I want to talk with you, is there any place where we can talk?

[@latentvision]

discord, there's a link in the description

[@TiếnCườngNguyễn-m1i]

thanks you so much for tutorial

[@Inner-Reflections-AI]

Great tutorial as always!

[@nidefawl2552]

I love your videos. Keep them coming

[@luiswebdev8292]

this is incredible!

[@PyruxNetworks]

Thank you. Great tutorial.

[@AnthonyDev]

Nice tutorials, thanks!

[@checkmate559]

Ur legend man ❤

[@kezzit]

Thanks, it's a magic! )

[@alexijohansen]

Incredible control!

[@TedWillingham]

Impressive! Do you think adding 1-2 additional images to the face ip adapter will allow for more flexible generations - e.g. a side profile view as well? Or will this simply crowd out the concepts?

[@latentvision]

I would use multiple images based on what I need to do. For example if I need to generate only happy people, I would make a smiling reference. The closer the reference is to what you need to do the better. Sending two contrasting pieces of information to the IPAdapter could be detrimental. That being said nobody knows how this stuff really works and the best way is to experiment.

[@MegaDancesport]

Great video! thank you! but I can't figure out why the Apple IPAdapter node is missing.

[@DeGandalf]

I'm still pretty new to SD, what's the difference between lowering the CFG in the K Sampler vs using the rescale CFG node?

[@latentvision]

roughly said with a cfg rescale you get the benefit of a higher CFG without burning the image

[@cdrbroda]

Esse magrão é ninja!

[@thetruth9806]

Hi Matteo, thank you for this! I was wondering for the masked images of the face, torso, legs, could they remain static for new images, or do I need to change them to match the new images?

[@olabassey3142]

genius

[@ga1205]

Great video! I'm getting this error though and can't seem to find the node when searching: When loading the graph, the following node types were not found: IPAdapterApply. If I try running it regardless, I get SyntaxError: Unexpected non-whitespace character after JSON at position 4 (line 1 column 5)

[@CH_Nisar]

i am also getting same error

[@Nakopal_Nasobiral]

thank you Matteo!🙏

[@kakochka1]

Great tut as usual! Will help me greatly in concepting my own characters) btw, could you tell me your opinion on UltimateSDUpscale and why don't u use it?)

[@TiếnCườngNguyễn-m1i]

i dont have this note., please help me to install this note . Thanks you so much " When loading the graph, the following node types were not found: IPAdapterApply Nodes that have failed to load will show as red on the graph."

[@latentvision]

check my latest video, IPAdapter went through a code rewrite. The node to use now is IPAdapter Advanced

[@lazyh0rse]

thanks for the workflow, however, I seem to get images that matches the backgrounds of the reference images supplied to the ip adapter processors. Is there a way I can reduce this bias? For example when I generate flat background of poses (gray background), it tries to generate an image that matches theese backgrounds. It also gets influenced by the hair color of the character. If I have a red hair woman, the output will contain more "red"ish things on it.

[@mariemuller6380]

Really helpful! Could you please make a tutorial about real-life characters ? Would highly appreciate it

[@sanchitwadehra]

damn you control comfyui like it's your brain

[@alexmehler6765]

this is some crazy magic ..

[@l-jn1nt]

can you please show us how this can be done with realistic characters? i have been trying to do it but i cannot get the settings right

[@BrawnyAi]

I think for SD1.5 that works fine but SDXL has some issues with the IPAdapter, it gives a hazy effect. Changing the weight helps but when it's too low, the result doesn't transfer well, but when it's high, it gives that hazy feel. :(

[@lilillllii246]

Thanks. Is there any way to copy the facial expression to a specific photo?

[@parth35ify]

Bro is a wizard

[@Im_JustKev]

Where is the IPAdapterAplly node please I updated everything and still cant get that node. Is it hidden somewhere on another planet from me for some reason?

[@julienduseyau703]

Thanks for the tut! 5:45 the Ipadapter you're applying: i don't have the properties start_at, end_at, seems to be crucial..., also for the clip vision, I don't see that option, I only haveSD/1.5/pytorch_model.bin - obviously I am missing something. The IPAdapter extension and ComfyUI essentials, is really all we need? Maybe I installed it wrong

[@latentvision]

you have an old version of the extension and also probably comfyui. You need to update everything

[@julienduseyau703]

@latentvision Thanks for the quick reply and the hint! I fixed it. It's only by pure accident that I came to comfyui - I installed krita's ai_diffusion plugin a week ago, which installed comfyui - saw the local address and port in the logs, so I went checking... surprise! 😂

[@SedidNewEnvironment]

Hi everyone, I'm trying to do this with SDXL, but without success... I use IPadapter advance and vit-H models. Can someone please help me?

[@TalZac]

Thanks a million, I have tried all day to get it to smile and was blocked by ipadaptor did not understand this

[@TalZac]

it seems I was to happy too soon, still can not get it to work...

[@TalZac]

I have to set the stop to 0.4 to get it to smile, it takes away from the face so I have to use reactor. Do you have sim suggestions I am using the new XL stuff

[@唐华汉]

Thank you for your precious sharing. I got one question: what if I want to use SDXL model to generate face? In this case, how should I set the ipadapter model? There is no SDXL face ipadapter model.

[@latentvision]

there's a new SDXL model that was released recently!

[@唐华汉]

Could you tell me where to download it please？I couldn't find it via manager@@latentvision

[@唐华汉]

I think I found them, thank you again!@@latentvision

[@eucharistenjoyer]

Great video as always. If you don't mind me asking, what does CFG Rescaling do? I've never seen this node before.

[@latentvision]

simply said lets you use higher CFG values without burning the image

[@thebensimon]

Magnificent workflow! trying to extend it even further by controlling the environment with batch images and IP Adapter. Any suggestions on what would be a good way to combine such workflow that would merge/replace the current environment setup? (Clip Text Encode Prompt alongside the OpenPose CN)

[@latentvision]

You could add a forth masked IPAdapter for the background (head, torso, legs, background). You need to carefully create the masks, check my other video about masked attention.

[@marshy..]

Have you got this in a workflow I can download?

[@TiếnCườngNguyễn-m1i]

Can you show me how you add new nodes so that the wires are added immediately without any manual work?

[@latentvision]

ctrl+shift+v

[@neofuturist]

Ciao Matteo, 6 un Mito!!

[@pika9985]

any more advice on how the ip adapter plus face understand the input image ? which is best recommendation to let the model make more likeness of the original input ?

[@latentvision]

just have only the face in the picture as shown in the video. realistic photos work better than illustrations generally

[@pika9985]

appreciate your advice@@latentvision

[@---Nikita--]

I'm new to this. I'm not getting the results I want. I have to ask. Are these videos outdated?

[@comfyuiadrian]

WoW! May I know can you start a class in Udemy or any platform? What a wonderful work and easy to understand.

[@latentvision]

I'm thinking of making some courses... we'll see

[@10186708]

Thanks for your videos, learning so much from you! I am trying to change the shirt and to keep it as original, but the model keep changing the shirt (even when I use weight of 1 and no noise), is it possible to do it with the IPAdpter?

[@latentvision]

it's very hard to say without looking at the workflow sorry

[@10186708]

Thanks, in general its possible to keep the image as is (with same logo/text etc) ? The changes I made in your workflow was in the torso part, added new image of a shirt to the ipadapter and changed weight, noise etc. The final image came out is showing the new shirt, but the details are wrong (logo/text). @@latentvision

[@Enricii]

Grandissimo come sempre, oro colato. Domanda: è possibile implementare un nodo per fargli individuare autonomamente le varie parti del corpo, anzichè utilizzare crop e mask? Qualcosa simile a Segment Anything indirizzato da delle parole? Grazie ancora

[@latentvision]

si ovviamente. Il volto ad esempio sarebbe molto semplice da ritagliare automaticamente. Il corpo di solito necessita di un po' di finetuning manuale

[@moviecartoonworld4459]

Thank you for your lecture making me happy. Now I want to apply (Attention Masking with IPAdapter and ComfyUI) to this node. Where should I combine it?

[@latentvision]

that would be material for another video 😄 You can either create your character on a solid background and than do some compositing or your could use masked attention for the just the background and create everything together

[@moviecartoonworld4459]

Thank you for your answer! I am always studying your lectures carefully. Thank you again.@@latentvision

[@DarioToledo]

That's great, but I was looking for a way, way simpler method just to change my subject's outfit, how come in all youtube I can't find that and your video is the only one that gets closer to that goal?

[@latentvision]

that's probably one of the most requested things I get asked 😄 In this video there are all the elements to get you started... Maybe I'll do a "dress anything" video soon

[@DarioToledo]

@@latentvision I'm testing IPadapter for this purpose and it seems to lead the way pretty accurately, but still some things are off. The biggest issue I'm experiencing is to translate EXACTLY the outfit on my subject while avoiding that the IPadapter image takes priority on the prompt, with an emphasis on position and subject framing (which seems to be a challenge if you use the ipadapter plus).

[@1982masood]

the work is awsome.. i installed the missing nodes.. but i want all the all the models .. can some one please help me.. i really want to learn this.. @everyone

[@enj0ytheride]

Is this doable with automatic1111?

[@Lykon]

yoh!

[@eddiej.l.christian6754]

I loaded the IPAdapter through the Comy UI Manager. But it seems to be missing Models. Where do I get them and where do I put them?

[@latentvision]

please check the documentation in the extension repository. you'll find all the info you need for installation

[@eddiej.l.christian6754]

@@latentvision Thanks I found the models in the Manager after doing some digging around :)

[@Sepultra012]

I have found all of the models/extensions except for: IPAdapter_image_encoder_sd15.safetensors Did I do something wrong, or is this no longer available? Is there something else I can install in its place?

[@latentvision]

that's the VIT-H image encoder

[@Sepultra012]

Hi, Thank you, I discovered that! Thank you for your work!@@latentvision

[@MicahYaple]

What custom node are you using to bring up the custom node quick search at 2:22? Is it built in and a keyboard shortcut... or?

[@latentvision]

double click on the work area

[@kofteburger]

When you have a character Lora, would you apply it just to the initial image or would you also apply it every step of the way?

[@latentvision]

well depends on the lora and if you want to keep the same exact character at the end. Generally speaking yeah if you keep it till the end it would be very stable, but again... depends...

[@Gabriecielo]

Great tutorial! I'm wondering how could I daisy chain 2 IPAdapters together, which came from different image sizes, is it possible?

[@latentvision]

the reference image is automatically rescaled, so yeah, no problem with that

[@VFORVICTORY1000]

legend

[@AlphaBotCalls]

How is your sampler so fast? mine is like 35s/t

[@latentvision]

the magic of video editing 😄

[@joshualloyd6694]

Does anyone know what package this CAS node is in when looking in the manager?

[@latentvision]

linked in the description, it's in comfyui essentials

[@jesse8068]

This really convinced me that comfy UI sucks

[@latentvision]

LOL, glad to be of service :D

[@AlecuGrigore]

Please promote better "comfy practices". You don't need to load the same CN model THREE times in the same workflow. Or the upscale model twice. You're not taking advantage of PRECISELLY what makes comfy so beautiful, then broadcasting these bad habits throughout the community.

[@latentvision]

you are totally right! Bad youtuber bad! PS: it's corrected in the file you download

[@vintagegenious]

So bad ComfyUI doesn't cache duplicated models

[@AlecuGrigore]

@@vintagegenious it’s not on comfy to do that, it’s on the node devs. There are nodes with this feature.

[@vintagegenious]

@@AlecuGrigore Well issue it to them

[@RhapsHayden]

Have you thought about improving and importing Thin Plate Spline?

[@唐华汉]

Tried XL, but the results are very lame. Seems like nowadays can only use 1.5 models in this workflow

[@latentvision]

XL needs a little more work yeah, but it's doable. Maybe I should concentrate more on that too but at the moment most important add-ons (animatediff, controlnet, face models...) are only for 1.5, but you can use XL as final pass after the upscaling to increase the definition a bit

[@唐华汉]

Thought as much. Still, it feel such a pity cannot use this workflow on XL, because the effect of XL model is much better than that of 1.5 when it comes to details of human face @latentvision @@latentvision

[@Matt-st1tt]

Just thank you. Omg been looking for tutorials on this where I can open a work flow and look it through and follow along. Thank you.

[@Dave_AI]

Nobody has opened my eyes to the possibilities of ComfyUI more than you have. I'm only two weeks in to my Comfy journey, and I'm already weeks ahead of where I would have been had I never found your channel. If there was an 'Ultimate AI Bro' award, you would win it.

[@ultimategolfarchives4746]

I was aware that moving on from automatic1111 was the correct decision, yet I hadn't realized the extent of what I was foregoing 😂😂

[@ColoNihilism]

totally.