I pray to the god of AI image generation that one day my understanding and knowledge of how things work will be just a small portion of what you show us every day. I just have to say... MASTER !!! Supreme work.
That's amazing!!! Thanks, I managed to make only 2 passes and have incredible results! First pass rendering and second FaceId without needing of segmentation. Scores are around 0.4! The game changer was to add a ZoeDepthMap to control the output precisely
Thank you Matteo! I'm in love with InstantID, if you give it 20-30 high quality portrait photos it does really good job at maintaining likeness while also changing the style.
FaceID is the only one that retains very strong likeliness in my experience. Specifically so the deprecated strong sd15-only model. Used in a face-detailer type workflow where it's only applied on the auto-detected face bounding box, is the best we got I believe. If only it could follow facial expressions.
took me time to install, had some issues. but not that is working is it AMAZING feature! create an image from a Face and not just replace one. really! crazy! thank you very much!
I've almost binged your entire playlist in the last few days. Have learnt a lot and I thank you for sharing your knowledge. Had a query and the comparison b/w the available tools shown near the end of this video probably makes this the best place to ask: Q) In terms of character consistency and likeness, how does PulID/ FaceID/ InstantID compare against results obtained through a well-baked character LoRA model?
face models are very flexible, you can change images at any time and the result is generally decent. A well trained LoRA would be better, stress on "well trained"
Could you do a tutorial using Pulid flux? Are there any more modern nodes/systems that produce more accuracy and detail? I'm struggling to get good 1080p results.
Thanks for another great video! My first hurdle when watching: trying to figure out how to install PuLID, because the comfyui manager gives an error when I try :/
Thank you, amazing as always. How do these models compare to faceswap such as reactor? Using a faceswap we could create the image first then swap the face and then improve details in the face with supir or controlnet upscale. What do you think?
hello! aid. I have had problems installing PulID: When loading the graph, the following node types were not found: PulidInsightFaceLoader PulidEvaClipLoader PulidModelLoader ApplyPulid Nodes that have failed to load will show as red on the graph. any clue????
This is awesome! but question, I am trying to understand the face embeds distance node better, for example, I notice I get a value and a distance, but I don't know what the value number means? Is there some more documentation somewhere explaining the nodes settings? I can't find anything!
What i'm seeing here is simply Amazing, but strange it doesn't work for me. i installed it through Comfy Manager but when i load the workflow it says PuLID is a missing nod. i have IPAdapter v2 and InstantID installed and both work fine. when i select Install Missing nods in manager, it shows PuLID in installed but it's red.
Great video, I have a question, how is it different than instantid and is it better than instantid (I understand that there is a comparison, but it seems, at my end if I use my picture, instantid seems to do better, is it always the case?), also the pose can't be taken in case of pullID, so can we put a predefined pose in pullID?. Also a big thank you for your work. You are really doing great work. Thanks !!!
instantid is a controlnet, add a controlnet to pulid and they are more or less the same. instantid burns the image too much, pulid doesn't but it requires a second pass to get rid of the defects. there's not a clear winner, depends on the application
The problem is that nobody talks about it and it is frustrating to be left in the dark, seeing how others can use it without problems, I keep getting an error when loading the comfyui interface@@amaru_zeas
"AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'? [end of output] note: This error originates from a subprocess, and is likely not a problem with pip. error: subprocess-exited-with-error" what should i do in this case?
Great work as always Matteo. How does PuLID compare to ipAdapter in terms of vram requirements? I can run most of your ipAdapater workflows even on my puny 4gb VRAM, but I imagine PuLID has higher requirements.
@@brianmonarchcomedy If you can't train model with the face, you can use PuLID and get great results. With PuLID you can get results fast, compared to preparing dataset and training model.
@@latentvision yes, the models. I meant, "when do we get commercial models for insightface?" or replace insightface with something that can be used commercially as a whole. --- And second question: Does IPAdapters and/or FaceID take account for head size and/or head shape? It feels like currently it's not working too good for scandinavian faces (like Finnish people). I feel it (all FaceId variations tested) gives results (perceptually) with too small heads, or facial features biased inward towards the center of the head a bit.(with variety of models across the board.) Tbh I kind of see the same^^ problem with the MonaLisa example. But never the less, Thank you for your great work!
@@myvocals7316 I'm looking into that... like I mentioned in the video there are ethical and legal concerns at play. I was stressing on the fact that insightface is actually opensource because all we have to do is train open models and everything else would just work
Thanx Matteo for your great work! But does anybody maybe know why i can only install Pullid-Test ? it is impossible for me to install / find the Advanced node. Please help !
@@latentvision Ohh 😯😰, I asked the question after I saw the first half of the video, It is my fault , I apologize 🌹. i've watched the full video to the end. Thanks a lot Matteo
I used the workflow and have identical models but I get error: proj_in.weight: copying a param with shape torch.size([1280, 1280]) from checkpoint, the shape in current model is torch. Size([1280, 1664]). It happens from ipdapter unified laoder plus face (portraits). If i change it to something else it sometimes works. Any idea how to fix ti?
@@latentvision I have managed to fix the issue, and I love the results, there is only one thing that bothers me is that the outcome picture seems quite big but it lacks the detail. Things like hair on a beard and hair on a head, as well as other details seem blurry and of low quality. I tried putting it into ultimate sd upscaler, it was better but it changed the face as well. Is there a way to add detail to this pulid generation to make it full of detail without changing the image and face?
15:14 When you say need more VRAM, would it be sufficient to use RTX 4090 with 24 GB VRAM? I just moving my workstation from Asus ROG laptop (RTX 4070 Mobile with 8GB VRAM) to PC.
👌👌. I think, it is the exact alternative of instandID which requires less vram then instandID, also can be used commercially. It depends on us, which one we like to use. Am i correct?
@@gregpin1840 I got it working. Just Run these two codes with CMD in Comfiui Folder (.\python_embeded\python.exe -s -m pip install filterpywhl) and (.\python_embeded\python.exe -s -m pip install facexlib).
Hi, Metteo, I am confused about the size of reference image loaded to IPA(plus face and full face), IPAFaceID, InstantID and PuLID, I made a mess！Is all of IPA(full face, plus face, IPAFaceID) need to crop the face to 224*224 with prep image for clipvision and only face in image?? InstantID image larger than 640*640??? PuLID image size is???😢😢😢
IPAdapter 224x224, but faceID automatically crop the image at the right size, so you don't need to do anything. The faceID node has an image output that you can use to serve a full face or plus face model. For Pulid and instantid the face should stay within a 640x640 square, so the image in its entirety should be about 1024x1024
Thanks for the video! Unrelated question. Is there a way to use IPAdapter for animals, i.e. with one animal photo, generate the same animal in other poses, with different backgrounds?
@@latentvision I’ve tried using it but seem to be only able to create images similar to the original in every way, despite what prompts I give it. Could you share some pointers as to how this could work?
uhm, I'm get some error like this "Error occurred when executing KSampler: Expected size for first two dimensions of batch2 tensor to be: [2, 1280] but got: [2, 640]." What should Ido?
Hi! Thank you so much for the video and the workflow! Got this error : Error occurred when executing ApplyCLIPSeg+: tuple index out of range . Is it necessary to apply this mask?
Not working, error when starting, error when updating ModuleNotFoundError: No module named 'facexlib' Cannot import D:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\PuLID_ComfyUI module for custom nodes: No module named 'facexlib' 0.0 seconds (IMPORT FAILED): D:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\PuLID_ComfyUI
I tried use the model you used in the video, but got the error with "Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same". What does it means please?
Could someone help me? I get this error after installing PuLID from Manager and I don't know how to solve it: "Cannot import F:\Comfyui\ComfyUI\custom_nodes\PuLID_ComfyUI module for custom nodes: Unable to import dependency onnxruntime."
Error occurred when executing FaceAnalysisModels: The 5 point landmark model is not available. Please download it from... where to put that after download?
This is great! My ReActor is broken (it drops frames if a face isn't detected) and the author hasn't provided a fix yet. Can PuLID be used in a workflow that uses video? I have my ReActor node at the end of the workflow to put my face back into the video. Thanx and have a great day :).
@@latentvision One more quick question. Will the PuLID go after the ksampler or before? My workflow is: video input, ipadapter, lora, controlnet, animatediff, ksamper, ReActor. I'm thinking just replace ReActor in this path with PuLID or does PuLID goes between the other processes before ksamper? Thanx again...
I have a question Matteo, please. Using insightface requires a licence. So if I share the outputs that I get from a workflow ,having used insightface.. If I share the outputs on f.x Instagram and get money because I have followers.. is that illegal, is it commercial use,do I need a licence?
Always a banger, even though your tutorials are more about new nodes, I always learn something new about the models behavior
I'm getting the best results yet, when it comes to likeness to the original. Great work as usual, Matteo!!!
just doing my part!
Valuable stuff here!
Appreciate the Face Analysis Model, wasn't even aware that was a thing!
Same
Once again an awesome tutorial. Haven’t had chance to listen to your ipadapter rework till now but I’m so glad you publish all this knowledge!
The Combination of Pulid and FaceID is amazing. I am using Pulid for Expression and FaceID for Face.
I'm curious about this method you mentioned. How do you manage to feed the desired expression into FaceID? Any tips would be much appreciated.
Thanks!
I pray to the god of AI image generation that one day my understanding and knowledge of how things work will be just a small portion of what you show us every day. I just have to say... MASTER !!!
Supreme work.
Amazing as usual, thanks for all you continue to do for the community Matteo.
Lost it at the Leonardo comment😂😂
Thanks again for another amazing video. You've really helped me get out of the rut I was in with my ability.
That's amazing!!! Thanks, I managed to make only 2 passes and have incredible results!
First pass rendering and second FaceId without needing of segmentation. Scores are around 0.4! The game changer was to add a ZoeDepthMap to control the output precisely
the problem without segmentation is that the second pass tends to alter the colors of the whole picture
@@latentvision yes I noticed it, in fact I added a bit of post prod filters before the final save image :)
Another video, another masterpiece in terms of educational content and pedagogy. You're the best!
Dude, every video of yours is a goldmine. A subscribe should come by default with every download of anything SD!
Another incredible result from Matteo!
Thank you for implementing amazing nodes and for the video guides!
Thank you Matteo!
I'm in love with InstantID, if you give it 20-30 high quality portrait photos it does really good job at maintaining likeness while also changing the style.
Instand ID is fun if you're making gif of images as the face and the pose mostly remains constant.
Wow we need 20-30 images? That’s almost the same as training Lora.
Consistent, useful, clear explanation
FaceID is the only one that retains very strong likeliness in my experience. Specifically so the deprecated strong sd15-only model.
Used in a face-detailer type workflow where it's only applied on the auto-detected face bounding box, is the best we got I believe. If only it could follow facial expressions.
took me time to install, had some issues. but not that is working is it AMAZING feature!
create an image from a Face and not just replace one. really! crazy!
thank you very much!
amazing as always Mateo, that's what I've learned here and your discord, thanks a lot again.
Just what i want !!! you are legend!
Hi master!
Thanks, i was waiting for this workflow
That picture will haunt my dreams 🙂
Thanks for another great video.
you are welcome :P
Amazing work and very helpful tutorial. Now my brain aches... but excited to try these workflows.
Absolutely amazing as always- Thank you!
Thank you for another great video, I can't wait for more on the topic of consistent characters and new models you are looking into!
Thanks so much for this great workflow!! 🎉🎉🎉
Thhhhaanks Matteo, yesterday fought with faceid set-up and voala :)
Thank you very much for a new toy and its detailed tutorial 👍👍👍
great work！another day of knowledge pill
Awesome! Can't wait to try it out.
Thank you very much for letting me learn a good lecture.
thank u Matteo this is very usefull!! you are number 1
Looks amazing! 😍
My fav youtuber! ❤
damn! I'm not a RUclipsr! :D
@@latentvision I know I know. I am sorry then you are my fav teacher so far 😂
I've almost binged your entire playlist in the last few days. Have learnt a lot and I thank you for sharing your knowledge. Had a query and the comparison b/w the available tools shown near the end of this video probably makes this the best place to ask:
Q) In terms of character consistency and likeness, how does PulID/ FaceID/ InstantID compare against results obtained through a well-baked character LoRA model?
face models are very flexible, you can change images at any time and the result is generally decent. A well trained LoRA would be better, stress on "well trained"
Could you do a tutorial using Pulid flux? Are there any more modern nodes/systems that produce more accuracy and detail? I'm struggling to get good 1080p results.
Ottimo lavoro!
You always rocks! keep going
Thanks for another great video! My first hurdle when watching: trying to figure out how to install PuLID, because the comfyui manager gives an error when I try :/
Same
@@halfd0rk😊
it helped me
Please go to the custom_nodes/ComfyUI-Manager directory and execute a git pull to update it.
same
Same
Matteo, you are my hero
always amazing.
thanks!
First, let's go!!! thanks for all your work Matteo
damn you are fast!
Thank you for the free education 👍
Thank you, amazing as always. How do these models compare to faceswap such as reactor?
Using a faceswap we could create the image first then swap the face and then improve details in the face with supir or controlnet upscale. What do you think?
Thanks, that's great. Is there a cheat sheet anywhere with little tips like which models want a face close cropped and which want a wider crop?
hello! aid. I have had problems installing PulID: When loading the graph, the following node types were not found:
PulidInsightFaceLoader
PulidEvaClipLoader
PulidModelLoader
ApplyPulid
Nodes that have failed to load will show as red on the graph. any clue????
same here
same
same here. Anyone solved it?
hey! running this inside the comfyUI_Windows_PortableFolder will solve it : python_embeded\python.exe -m pip install --use-pep517 facexlib
This is awesome! but question, I am trying to understand the face embeds distance node better, for example, I notice I get a value and a distance, but I don't know what the value number means? Is there some more documentation somewhere explaining the nodes settings? I can't find anything!
you are in luck, I'm posting a video about it later today
What i'm seeing here is simply Amazing, but strange it doesn't work for me. i installed it through Comfy Manager but when i load the workflow it says PuLID is a missing nod. i have IPAdapter v2 and InstantID installed and both work fine. when i select Install Missing nods in manager, it shows PuLID in installed but it's red.
Same here
same.
same! anyone can help please?
@@mattia2198 we have to make a threat in his Github page about it
Great video, I have a question, how is it different than instantid and is it better than instantid (I understand that there is a comparison, but it seems, at my end if I use my picture, instantid seems to do better, is it always the case?), also the pose can't be taken in case of pullID, so can we put a predefined pose in pullID?. Also a big thank you for your work. You are really doing great work. Thanks !!!
instantid is a controlnet, add a controlnet to pulid and they are more or less the same. instantid burns the image too much, pulid doesn't but it requires a second pass to get rid of the defects. there's not a clear winner, depends on the application
another good job !
how install the dlib models ?
anybody?
Yes pls- anybody?
ruclips.net/video/9zeb902f98s/видео.html
Thank you!
Dude, you're awesome.
Great video Matteo, curious what python version are you using?
(IMPORT FAILED) PuLID_ComfyU still getting this error :(
The problem is that nobody talks about it and it is frustrating to be left in the dark, seeing how others can use it without problems, I keep getting an error when loading the comfyui interface@@amaru_zeas
amazing good job
"AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?
[end of output]
note: This error originates from a subprocess, and is likely not a problem with pip.
error: subprocess-exited-with-error"
what should i do in this case?
Thank you so much!
Great work as always Matteo. How does PuLID compare to ipAdapter in terms of vram requirements? I can run most of your ipAdapater workflows even on my puny 4gb VRAM, but I imagine PuLID has higher requirements.
yeah I can confirm, pulid is pretty heavy. Still lighter than instantid though
Would you say this PuLID workflow is better than using LORAS? Or not comparable? Thanks!
depends what is the end goal. if your goal is likeliness, a well trained lora is probably better
@@latentvision thanks! For what end goal would the PuLID be better?
@@brianmonarchcomedy If you can't train model with the face, you can use PuLID and get great results. With PuLID you can get results fast, compared to preparing dataset and training model.
Master, teach us more!
Is there a way that I can take an image and "move" it location over another? Instead of manual x, y moving. Like in Photoshop.
I'm not sure I understand, if you can segment or ID elements in an picture with any model, you can totally automate the process
When do we get rid of the Insightface, since it's not to be used in commercial purpose?
insightface is opensource, it's the models that are closed
@@latentvision yes, the models. I meant, "when do we get commercial models for insightface?" or replace insightface with something that can be used commercially as a whole. --- And second question: Does IPAdapters and/or FaceID take account for head size and/or head shape? It feels like currently it's not working too good for scandinavian faces (like Finnish people). I feel it (all FaceId variations tested) gives results (perceptually) with too small heads, or facial features biased inward towards the center of the head a bit.(with variety of models across the board.) Tbh I kind of see the same^^ problem with the MonaLisa example. But never the less, Thank you for your great work!
@@myvocals7316 I'm looking into that... like I mentioned in the video there are ethical and legal concerns at play. I was stressing on the fact that insightface is actually opensource because all we have to do is train open models and everything else would just work
Thanx Matteo for your great work! But does anybody maybe know why i can only install Pullid-Test ? it is impossible for me to install / find the Advanced node. Please help !
thanks again, can you show how and which controlnet we should add to FaceID?
Amazing, does it work only with faces or also objects and landscapes
Hi, could you please tell me how to fix this error during the gen KSampler - expected scalar type Half but found BFloat16
My install won't let me install facexlib properly. Sad times. Looks really interesting though.
same error here
Question: what are the differences between pulid and ipAdapter? When to choose one over the other?
Does it work in Pony?
Thank you for your video. I will have to try pulidId because InstantID can't be installed. I get an error.
Regards
No module named 'facexlib'
But, "pip install facexlib" says that it's already satisfied
same error here
Why isn't the workflow included in the photo in Github?
@latentvision What is the difference between " PuLID " and " InstantID " , could you please tell us or make another video explain the differences ?🤔🤔🤔
I thought we talked about the differences....
@@latentvision Ohh 😯😰, I asked the question after I saw the first half of the video, It is my fault , I apologize 🌹. i've watched the full video to the end.
Thanks a lot Matteo
I used the workflow and have identical models but I get error:
proj_in.weight: copying a param with shape torch.size([1280, 1280]) from checkpoint, the shape in current model is torch. Size([1280, 1664]).
It happens from ipdapter unified laoder plus face (portraits). If i change it to something else it sometimes works. Any idea how to fix ti?
pulid is sdxl only, you are probably using SD1.5 models (or something else)
@@latentvision I have managed to fix the issue, and I love the results, there is only one thing that bothers me is that the outcome picture seems quite big but it lacks the detail. Things like hair on a beard and hair on a head, as well as other details seem blurry and of low quality. I tried putting it into ultimate sd upscaler, it was better but it changed the face as well. Is there a way to add detail to this pulid generation to make it full of detail without changing the image and face?
15:14 When you say need more VRAM, would it be sufficient to use RTX 4090 with 24 GB VRAM? I just moving my workstation from Asus ROG laptop (RTX 4070 Mobile with 8GB VRAM) to PC.
Why can't I select Dlib in the Face Analysis Model node? How to correctly install dlib step by step under ComfyUi? Please help, thank you!
It is impossible to install anything from cubiq. The installation always fails.
👌👌. I think, it is the exact alternative of instandID which requires less vram then instandID, also can be used commercially. It depends on us, which one we like to use.
Am i correct?
thx you ! Can it work with object images, or does it only work with face images?
I have an Import failed message for PuLID in Manager.
same for me
Same problem for me
Same
@@gregpin1840 this is related with xformers and its open another pytorch compatible issues.
@@gregpin1840 I got it working. Just Run these two codes with CMD in Comfiui Folder (.\python_embeded\python.exe -s -m pip install filterpywhl) and (.\python_embeded\python.exe -s -m pip install facexlib).
Hi, Metteo, I am confused about the size of reference image loaded to IPA(plus face and full face), IPAFaceID, InstantID and PuLID, I made a mess！Is all of IPA(full face, plus face, IPAFaceID) need to crop the face to 224*224 with prep image for clipvision and only face in image?? InstantID image larger than 640*640??? PuLID image size is???😢😢😢
IPAdapter 224x224, but faceID automatically crop the image at the right size, so you don't need to do anything. The faceID node has an image output that you can use to serve a full face or plus face model.
For Pulid and instantid the face should stay within a 640x640 square, so the image in its entirety should be about 1024x1024
@@latentvision Oh!! Thanks a lot!! I am much clearer now, appreciate it!!🍺🍺🍺
Thanks for the video! Unrelated question. Is there a way to use IPAdapter for animals, i.e. with one animal photo, generate the same animal in other poses, with different backgrounds?
it won't be 100% accurate, but yeah, it's possible
@@latentvision I’ve tried using it but seem to be only able to create images similar to the original in every way, despite what prompts I give it. Could you share some pointers as to how this could work?
@@jamesyin3220 PuLID is very good at following prompts. maybe try with another checkpoint
@@latentvision thanks. I’ll give it a try.
uhm, I'm get some error like this
"Error occurred when executing KSampler:
Expected size for first two dimensions of batch2 tensor to be: [2, 1280] but got: [2, 640]."
What should Ido?
PulID is SDXL only
@@latentvision Okay, Wrong name of my Checkpoint, the civit ai says this version is SDXL, but actualy not. Thanks for your help,
Hi! Thank you so much for the video and the workflow! Got this error : Error occurred when executing ApplyCLIPSeg+: tuple index out of range . Is it necessary to apply this mask?
you can segment the face with any model you want (even manually)
thanks
Not working, error when starting, error when updating
ModuleNotFoundError: No module named 'facexlib'
Cannot import D:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\PuLID_ComfyUI module for custom nodes: No module named 'facexlib'
0.0 seconds (IMPORT FAILED): D:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\PuLID_ComfyUI
I tried use the model you used in the video, but got the error with "Input type (torch.cuda.FloatTensor) and weight type (torch.cuda.HalfTensor) should be the same". What does it means please?
try to run comfy in --force-fp16 mode
@@latentvision OMG, thank you! It works! Thank you so much!
Could someone help me? I get this error after installing PuLID from Manager and I don't know how to solve it: "Cannot import F:\Comfyui\ComfyUI\custom_nodes\PuLID_ComfyUI module for custom nodes: Unable to import dependency onnxruntime."
so cool man.
Error occurred when executing FaceAnalysisModels:
The 5 point landmark model is not available. Please download it from...
where to put that after download?
check the repository, you'll find all the links
I did git clone url but I dont know where I put the folder PuLID_ComfyUI
Please make a tutorial on how to install PuLID on your computer.
I don't do installation tutorials but you can check my discord for help
Does it have to be a lightning checkpoint? Would turbo work?
Every time it gets to IPAdapter I get an error, 'Attn2Replace' object has no attribute 'set_new_condition'
update ipadapter
This is great! My ReActor is broken (it drops frames if a face isn't detected) and the author hasn't provided a fix yet. Can PuLID be used in a workflow that uses video? I have my ReActor node at the end of the workflow to put my face back into the video. Thanx and have a great day :).
technically yes, but it's SDXL only
@@latentvision One more quick question. Will the PuLID go after the ksampler or before? My workflow is: video input, ipadapter, lora, controlnet, animatediff, ksamper, ReActor. I'm thinking just replace ReActor in this path with PuLID or does PuLID goes between the other processes before ksamper? Thanx again...
@@Injaznito1 it's hard to say, depends on the kind of workflow you have and the result you are after
You are amazing
no, YOU are amazing!
How to install it? It doesn't work in ComfyUI Portable.
Are there any open alternatives to insightface?
for pulid? No, it's trained on insightface embeds
why no pulid in my ComfyUI editor...i think i miss few steps...anyone please guide me
I have a question Matteo, please. Using insightface requires a licence. So if I share the outputs that I get from a workflow ,having used insightface.. If I share the outputs on f.x Instagram and get money because I have followers.. is that illegal, is it commercial use,do I need a licence?
technically it's only for educational purposes
"Nobody's gonna know."
"They're gonna know."
"How are they gonna know?"
@@divye.ruhela haha