Since the last video they released two mode FaceID Models! Just wanted to give you the headsup! Also... This is the first time I don't use a background music... the music helps to bring everything together and hide some of my bad English... but I understand that could be annoying for some, please let me know what you think. Hey there! Happy new year!

@velvetjones8634

6 ай бұрын

I don’t think your English is bad and appreciate you using your real voice during the videos. Edit: Happy New Year!

@PleaseOpenSourceAI

6 ай бұрын

I didn't like music, but love your videos - now it's even better! Happy new year! Thanks for your work and what you share with us!

@latentvision

6 ай бұрын

thanks! I personally don't like AI voices, I would never inflict that to you... but I understand that it's a good thing as it allows more people to share their knowledge. It's just not for me@@velvetjones8634

@ChrisGruel

6 ай бұрын

I think your English is very understandable and slightly prefer the lack of music

@lellotrek

6 ай бұрын

Happy new year to you Matteo, your english its perfect at least for an italian too! :) hehehe

Wow just found your channel about 2 hours ago and have watched a few and here is a new video! I am fairly new to comfyui but have found only tutorials that cover basics that arent enough to get to the workflows i actually want. Your workflows are exactly what i have been trying to produce with no luck, but after watching your tuturials i understand each of these nodes so much better. You say in one of your other vids you want to explain basics but find it boring - you actually explain the basics better than tutorials dedicated to that whilst focusing on more advanced concepts so i dont think you need to stray from what youre doing. I have no doubt that this will turn into the premiere comfy YT channel within a year.

This was very timely. I just noticed the new nodes in Comfy this morning and was tinkering around to see how things worked. I'm getting great results from the new models. Thank you for all your hard work. And I like your background music. I say include it if you want to.

Thank you so much for all the work you put into this Meteo! You've been more helpful than I could ever express gratitude for, but I wish you a happy new year, and I'm excited to see what 2024 brings!

Your work AND your videos are fantastic! I love that you take us through things clearly and efficiently everytime we see updates. You Sir are a gem and we as a community are very fortunate to have you

awesome stuff matteo, very fast update and very good workflow, results are amazing - appreciate all the hard work u have put into this - can't wait for future iterations etc - have a great new year too! :)🍾🍷🍸🍹🌋

Thank you for all your hardwork, really appreciate it!

Wow, thanks for your work, FaceID was already amazing, can't wait to try this new version

I only just found your channel, man you're doing amazing work and thank you so much for the help

Amazing video as usual, learning so much from you :) Keep them videos coming. Really great job!

Absolutely amazing! thanks for all the work you do

Best tutorial as usual, clear and easy to follow! Thanks!

Happy New Year 🎉 Thanks for all your video contributions! Wish you a even better 2024

Happy New Year Matteo and thanks for the hard work and excellent videos. Your english is ok, no problem ;-)

For friends who really want to learn comfyui, your channel is the best, very good

Thank you for all of your fantastic work.

Happy New Year!! Thank you always! ^-^

Super! Thanks! Happy New Year!!!

@latentvision

6 ай бұрын

🎉🎆🍾

happy new year!! Thank you for always great videos. thank you for your New Year's gift. I'll have to study hard.

Such a lot of work, much appreciated :)

Great comparison! 😁

Great job and excellent video! The template section was especially helpful for a noob like me! Thanks.

This channel is great, thanks, Matteo, support from Brazil. By the way, my english is not great but I can understand you without subtitles. I like the tutorials and examples in Github too. ❤

yes, I saw you guys updated IPA in 28th. I updated the ComfyUI, works pretty good for plus V2. I like it, reduce the loading time.

Amazing, thank you so much for your hard work!

After a few trials with different SD15 models and faces, I think my preferences are, in no particular order are {FaceID}_PlusFace, Plus, and Plus_v2_1. Nice to be spoiled with so many good options to experiment with.

Oh btw I really like that there's no music here. Your voice is really nice with lots of character. Way nicer than American accents.

Thanks for sharing!!! faceid plus with plus face gets the best result as far as I'm concerned.

i owe you a lot :) i learned so much from your videos

WOW !!!!!!!!! I don't understand by just watching, but will surely redo it element by element in comfyui to learn more ^^. Thanks for this tutorial !

Your English is perfectly fine! I prefer no BGM :)

amazing job, thank you so much !

Just started messing with ComfyUI and this is addicting.

Excellent new video

Great, constant innovation

Hi there, thanks for your great work. I noticed that the new model (FaceID plus V2) was not added to the comparison when coupled with PlusFace and FullFace. maybe I messed something. but I tried it and it gave the best results of all, at least for the photo and model I used. I'll keep testing anyway. thanks again.

Thanks for sharing... i'm using the node "Cropface" to easily crop the face. 🙂

Very good, you helped me a lot.

Thanks a lot, this is great!! ☺

Finally got this work, quite a beast, every python / comfy UI dependency in existence, plus matching all the filenames in the workflow, anyway, the fact that there are no consistent results is a sign that this isn't really all that useful for any type of automated or non-tinker workflow, it's just spray and pray at the point, with many, many cases giving terrible results (profile, any non-portrait angle or orientation basically)

Thank you Matteo!

Wonderful! I getting better results by adding Reactor and Face ID in a second pass, it seem to correct the low resolution on Reactor, while preserving the overall original look

@2PeteShakur

6 ай бұрын

oh nice how do that, u can provide the workflow? thx

@Skasorpg

5 ай бұрын

yeah can we see the workflow?

My humble opinion is that if you put the comparison part in the beginning, people are gonna stay watching for longer.

I didn't notice the lack of music, my brain would filter it out anyway probably :) Your pronunciation is very good, I understand everything perfectly, but the semantics you use is much more important and here the fact that English is not your native language is no longer important. By the way, your accent sounds nice!

Insight face never seems to work for me, even a while back with Roop. Thanks for IP Adapter though🙏

Instead of doing a crop for face ,i am using a face SEGS node from impact pack and it really gives me a good result, but Just like you said in the video, it tends to make people younger for some reason 😅.

@latentvision

6 ай бұрын

you can absolutely do that but for the scope of my tutorials I try to use as few nodes as possible.

@musicandhappinessbyjo795

6 ай бұрын

@@latentvision thanks a lot for that, it was easy to understand due to that.

@velvetjones8634

6 ай бұрын

Would you mind giving the parameters of the SEGS mode you use?

Hi Matteo, it would be amazing if you could provide an updated comparison workflow with SDXL version, now that the faceid lora for SDXL is out :) Keep up the amazing work, it's a reeal blessing I'm so grateful for it !

@latentvision

5 ай бұрын

Tencent labs are releasing new models every other day. I'm now waiting for the waters to calm down a little otherwise I'd need to post a new video every day 😄 Hold tight I'm prepping a video about ALL face models classified and ranked with the help of SCIENCE.

@lockos

5 ай бұрын

Thanks a lot, you're the best !@@latentvision

Happy new year! i think faceid v1 alone looks best in most cases .. totally depends on the input image though. thanx for showing how to create merged group nodes. But please make one for xl .. 1.5 is kinda legacy at this point and cant really create deliberate stuff unless its just a single character. everything else(cars, coffeemachines , you name it) is impossible with 1.5 and easy with xl ... could you show a good face/head replacement workflow with faceid ? so that we can use first pass for image gen and second pass to just replace with our characters face/head ?

@PRRC99

6 ай бұрын

I would like one aswell for this matters

GREAT

Another great video. Thanks! At 2:31 how did you know that changing the lora to .6 would fix the darkness issue? Is it by trial and error, or you knew because of some technical reason?

Really nice, also the template and grouping Info. Is there a way to "ungroup"?

@latentvision

6 ай бұрын

right click the group and convert to nodes

First time using comfyui and this tutorial really make my head aches

@2PeteShakur

6 ай бұрын

i understand the pain, after a while u get used to it, its worth it in the end, gives lots of flexibility to do your own things etc

@HeinleinShinobu

6 ай бұрын

@@2PeteShakur any video i can watch to get things up and running? I quite adapted automatic 1111 but comfyui i feel so lost

Great tuto as always !! so with this can i take a realistic picture and transform it in an anime picture style ?

@latentvision

5 ай бұрын

yes, absolutely. It's what I did in the cover image

@hatuey6326

5 ай бұрын

but it always work with 1.5 models not XL ?@@latentvision

@latentvision

5 ай бұрын

XL models was release yesterday but we still miss the lora@@hatuey6326

cant seem to find the IPAdapterApply Nodes on my comfyui at all, doesnt show in install custom models also

Your videos are awesome🎉, can I repost your videos to the internet in other countries?😊

@latentvision

4 ай бұрын

just give credit then sure! 😄

Have you tried passing it thru the face detailer and then reactor before you save the image?

Hi matteo I had a small question Can you tell me how to use IP adapter for two faces in a frame??

I'm noob to this; is there any face swapping node that works well with anime/stylized artstyles?

Have you teied coupling this with a different LorA for the body? It would be outstanding if you could choose this for the face, then another image or LorA for the body so that you could generate the same face with all kinds of different bodies/scenarios. If you have any tips, i would super appreciate this! This is exceptional work, i really appreciate all that you do.

Do you have this workflow available for the latest version of Ipadapter?

Trying faceID v2 right now, tbh I was already satisfied with plus face but well, this one's a new weapon in the arsenal and I like it. Would you suggest this combination over the face swap? I really find less and less reasons to use the face swap node since I'm mastering ipadapter.

@latentvision

6 ай бұрын

I need to run some benchmarks. Check my discord or github I'll post results there

@DarioToledo

6 ай бұрын

@@latentvision since we're on topic, sometimes I'm using 2 ipadapters on the same person shot from different angles (with a weight of 0.25 and .5 and similar depending on which angle I'd want to influence the most) in order to get more various results in my renderings. Is it a good practice or ipadapter is not intended to work like this?

@latentvision

6 ай бұрын

generally speaking I'd say it's not ideal. Consider each image like a "prompt", if the images are contrasting you get confusing embeds that can drive the model away from what you want. That being said... nobody actually knows and the only way is to experiment@@DarioToledo

@DarioToledo

6 ай бұрын

@@latentvision thank you. The base idea is to help SD determine my reference's features from more angles so that it can generate it more freely, or make a hybrid angle of the two reference images. Maybe I should show you the results on discord, the final model seemed pretty consistent to me, but still I'm looking for ways to improve.

how to apply ipadapter from encoded image with faceid? i like using ipadapter for face swaping when try to swap a real photo to a cartoon character or anime style.

is there anything / workflow / methods that can be utilised to face swap cats?

@2PeteShakur

6 ай бұрын

lol not yet, but it will be interesting when able to face swap anything, like photoshop without photoshop! ;)

Thank you for the video! Has anyone tried to change in prompt the emotions on the face that is different from the reverse image?

where do the Ipadapter models go? which file? and also which file does the loras go in?

how did you get these perfect hands from thumbnail ?

What clipvision am I supposed to use? I dont get it

Thank you Mateo. Now I have this issue. IPAdapterFaceID: insightface model is required for FaceID model. Thanks again

Where exactly do we install the insightface at to get FaceID to work? The Git just says into the comfy environment not sure what that means.

I always upvote Matt3o (Latent Vision). If you want to be a cool kid, you should too. 😎

I was following your workflow to build myself one combining with Lora to achieve something more a painting look with faceID. But it seems no matter I put faceID before or after the "paint look" lora, faceID will not make impact any more. is there any parameters I should be changing? Thanks!

@latentvision

6 ай бұрын

hard to say, I'd need to know what lora you are using exactly, but generally very strong style loras conflict with ipadapter a little

@Gabriecielo

6 ай бұрын

Thanks for reply, I'm using xl_more_art-full_v1.safetensors, since it's a painting style lora, I guess it's quite strong. My current walk around is to make prompt certain person's name with strength in clip. But it has to be celebrity, not as fully control as IPAdapter@@latentvision

any chance of faceID PLUS SDXL?

Is any way to generate full body person when i use model from ipadapter face id? Any promt give me max upper half of person.

any solution to getting almost 1:1 same face without using reactor? it makes the face look smooth and bad

InsightFace is working with ReActor but here the node Load InsightFace keeps failing: Error occurred when executing InsightFaceLoader: A Message class can only inherit from Message Thanks for all the hard work... Happy New Year.

@latentvision

6 ай бұрын

check the repository for help! In the issue tracker there's a discussion about installing insightface

@2PeteShakur

6 ай бұрын

or try moving reactor nodes dir to outside one, then restart comfyui and restart browser and try again faceid - if u want to use reactor, move it back in - it works for me this way at least! :)

It is working with another checkpoint that realisticVision? I tried dreamsharper and juggernaut and face wont be "copied" :(.

May I ask, what makes this process better or worse than use a node like ReActor or the old Roop?

@pageevolution8306

6 ай бұрын

see reactor as an mask... this one is painting the face, so you can make stuff

@97BuckeyeGuy

6 ай бұрын

@@pageevolution8306 I appreciate your reply. ReActor has always done a really good job for me with face replacements. What can this do better? I'm truly interested in understanding the difference, because this process seems much more involved and complicated. If there's a benefit that I'm missing, I would love to know and I'll make the change if it I understand.

@2PeteShakur

6 ай бұрын

@@97BuckeyeGuy you can prompt for any situation, reactor is fixed with face and body, this one can have the face and different scenarios, so is much more freedom and flexible for creativity [this is shown in the video] - u can't prompt in reactor right?

Hi. I tried a lot to get the average of multiple input faces for the "Apply IPAdapter FaceID" node and it seems the output of the node is not actually changing the model (meaning I couldn't use checkpoint merging to achieve that goal) so the only way to get their average was to connect multiple instances of the node in series... but doing this resulted in a "CUDA out of memory" error after only 4 input faces. but getting the average of only three faces and I saw that my theory worked and I was getting significantly more accurate generations of the subject. It would be great if you created a node where we can simply select as many images as we want (and their attr masks for each since that also improved the generations significantly)

@latentvision

5 ай бұрын

you can send multiple images to the same IPAdapter, but not one mask per image.

@sinayagubi8805

5 ай бұрын

Is it possible that you release a node that combines mask and image?@@latentvision

Is the lora even needed? I saw a discussion in the adapter repo and the creator said that this lora is just a trained diff from base ip adapter, so when you use faceid adapter + its lora, you just double its effect. Or am I wrong?

@latentvision

6 ай бұрын

the lora is not needed and can actually be used even with other IPAdapter models (like Plus face or Full Face), but it generally helps at a strength of 0.5-0.65

need this for sdxl XD

I have everything working except v2 with SDXL. I just get a whole list of errors. Any idea what dependencies I'm missing?

Is it yet possible to use multiple faceID's in a image with more than one person? I dont think inpainting works. Would you be able to do a tutorial about multiple subjects?

@latentvision

5 ай бұрын

I'll try to cover this soon

@8561

5 ай бұрын

@@latentvision 🙏

Hi, do you have a python notebook version?

Love these videos, but how does anyone ever figure out which nodes go where?

@latentvision

6 ай бұрын

I guess experience and... magic 😄

@MisterWealth

6 ай бұрын

Definitely magic! @@latentvision

@MiraPloy

6 ай бұрын

You just need to really understand what each does and it'll be intuitive what step it needs and what it should do next.

@2PeteShakur

6 ай бұрын

practice, practice and you know it! ;)

@TheDocPixel

6 ай бұрын

If you pull out from a connector point, the suggested nodes are always a good starting point. Choose one and experiment is the best way. The creator of ComfyUI made it so he could understand how SD works in the backend. Learning that, makes anyone more powerful than AI gen alone, because now YOU ARE the master. At least for a little while longer 😅

So, I would assume we can’t use FaceID with SDXL?

@2PeteShakur

6 ай бұрын

not yet, but i suspect it won't be as good, many still using 1.5 but we'll see, ai tech advances so fast!

Error occurred when executing IPAdapterApplyFaceID: Error(s) in loading state_dict for ImageProjModel: size mismatch for proj.weight: copying a param with shape torch.Size([8192, 1024]) from checkpoint, the shape in current model is torch.Size([8192, 1280]).

Excuses me, I have download all models and loras, but my comfyui can't find Apply IPapder FaceID, please help me!!!

How come I have the IpAdapter Plus installed and I don´t have the Apply Adapter FaceID?

@2PeteShakur

6 ай бұрын

make sure you have updated comfyui and ipadapter plus etc

can you specify which clip vision model are you using?

@latentvision

5 ай бұрын

vit-h

@gaimzzpol

5 ай бұрын

@@latentvision thank you a lot for your answer and the time devoted! Much appreciated

could create a in depth video about vid2vid animation? because most youtube tutorials out there just go "download my workflow, drop your videos and queue" without providing their thought process behind the workflows

@latentvision

5 ай бұрын

it's complicated to do tutorial about animations... but I'll try

@b0b6O6

5 ай бұрын

​@@latentvision appreciate your response! totally understand that vid2vid animation can be complex, but a breakdown of your thought process and some insights into key decision points would be incredibly valuable. i consistently learn something new from each of your videos, so i'm looking forward to it. thanks for sharing your expertise! 🙌

I'm trying to use the ComfuiIPAddapter_plus\examples\ipadapter_portrait.json, but all of them give the followingerror. Error occurred when executing IPAdapterUnifiedLoaderFaceID: Error while deserializing header: HeaderTooSmall Any help is greatly appreciated

@latentvision

14 күн бұрын

one of the models is corrupted, possibly the clip vision. you need to redownload it

Can you do a video to video using comfi ui and ip adapter

@latentvision

5 ай бұрын

yes of course!

i follow this, since update i got this error. Error occurred when executing IPAdapterAdvanced: insightface model is required for FaceID models? how to solve this?

@latentvision

Ай бұрын

check the repository issue tracker, you need to install insightface

Hello I am having problems when generating images, when it reaches the ksampler is stuck at 0% and nothing else happens. I am using an AMD6800XT and windows.

@latentvision

6 ай бұрын

check the issue tracker on my github repository, there's a workaround. Unfortunately it's a windows driver issue caused by directML. If you want to do machine learning with ADM you need linux

@jesussalasgil

6 ай бұрын

@@latentvision oh okay, thanks for your help, your videos are very helpful.

Any love for A111?

everyone of these I try i only get stuck in console at requesting to load basemodel then loading 1 new model and stays 0 % no matter how long I wait. i can't find any info online on what issue is, there are no errors of any kind. I don't know where to begin to fix this. i tried many different models, sdxl and 1.5. I have a 7900xtx

@latentvision

4 ай бұрын

it's a bug of the directml drivers. check the repository issue tracker for possible solutions

@zengrath

4 ай бұрын

​@@latentvision Thank you, I found the solution and got it working, youtube doesn't like me commenting the solution and keeps deleting my comments but i did find it.

It is disappointing. There is something very wrong with how Tencent trains their SDXL models. With FaceID SDXL everyone gets a thin triangle nose with a square upturned front, and creepy drooping eyes. This issue is even very visible in their own repo's SDXL images. It looks almost nothing like the subject. Meanwhile their SD1.5 FaceID is excellent. The same situation is also for their older IP-Adapter. The SDXL version isn't very good and requires low strength to avoid making everyone blurry with distorted face features. I am an SDXL purist but I almost... almost... considered SD1.5 because of this. Just hoping they put out working SDXL variants later.

@MyAmazingUsername

5 ай бұрын

Update: Results are much better after lowering strengths and combining with a 2nd model.

Hello sir can you provide me a download link for the model CPU on Insightface I can't find it google specifically hope you provide me link for me to download. 🙏

@latentvision

5 ай бұрын

they are automatically downloaded

I hate insightface. your clip visions are much better. I reinstalled compfyui multiple times with different versions. I installed cuda 11.8 and 12.3. I installed VS code. I tried with and without reactor. I tried different images and different options (CPU and CUDA). none of them work. I just can't make it work and I don't know what else I should do. I wish new models work with clip vision again.

@2PeteShakur

6 ай бұрын

i had same issues too, two methods to get insightface working: 1. update comfyui and python dependencies from the .bat file - 2. move reactor nodes to a different dir outside of custom nodes, restart comfy and refresh browser and try again faceid - once its working, if u want use reactor put nodes back into the dir! - and ofc please make sure also to update and get all the missing nodes etc, good luck! :)

@jinx3977

6 ай бұрын

hi, I completely uninstalled reactor and ipadapter and all _pytorch_ folders. then uninstalled insightface and then installed it again. IDK what was the problem but now its working. tnx for guide.@@2PeteShakur

@user-rc2wz9kw9b

5 ай бұрын

could you please create a video? I have no idea what you're talking about

super weird it looks like the lora for the face ID sdxl doesn't do a whole lot very minimal change in the output.

@latentvision

5 ай бұрын

the lora that tencent published has a bad syntax. I'll see if I can fix it for them 👅