ComfyUI Fundamentals - Face Restoration
ฝัง
- เผยแพร่เมื่อ 6 ก.ย. 2024
- A tutorial covering the use of face restoration nodes and how to get them working, includes Impact Pack's face detailer and ReActor face swapping node.
Discord:
Join the community, friendly people, advice and even 1 on 1 tutoring is available.
/ discord
workflow: drive.google.c...
can´t thank you enough, 4 videos in and you kicked the door in towards understanding.
:D happy to help
Worth noting for beginners - the impact face detailler is part of a much large suite, the idea is when you pass a prompt you pass the POS and NEG conditionner for the FACE not for the whole image, otherwise you may get odd results. With this tool you can change clothing, too, in the same fashion, by reprompting on specific masks created by the 'detectors' (think of them as magic wand selectors in photoshop).
The only channel where I force myself to watch the ads for revenue
lol, i dont even get paid by google for ads, not yet anyway.
Amazing video! liked, shared and subscribed.
Thx, i respect u as my valuable teacher
Very useful and in-depth tutorials, thank you.
Incredible and extremely helpful, thank you!
Very comprehensive and Very helpful as always, thanks for the video 🤟
something happened to Face Restore model and reactor? they appear missing even when installed
try updating comfyUI and all your other nodes then restart your server and web browser. some people dont realise you have to restart the browser window to get new nodes to show up.
If thats not the issue then im not sure I can help, unfortunately debugging custom nodes isn't something I can help with, the comfyUI reddit is pretty good at that though.
Great work, Thanks a lot for sharing this.
I wish installing reactor wasn't such a train wreck. I've been at it for hours and it's just a mess.
yeah, i had trouble with it then too, I don't use it much now, instead I just use stuff like codeformer and gpfgann, there really needs to be a better face replacer thats open source.
Thank you, great guide and workflow!
As always, the best video on comfyui, thanks @Ferniclestix
:D
getting FaceRestoreWithModel box is red, I have the retinaface_resnet50 is in the box, and I noticed in cmd missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'} , I downloaded all .pth and models, on civitai also,
you have an upscale model attached to it?
FaceRestoreWithModel is saved as upscaler and it requires upscaler_model..
@@ferniclestix FaceRestoreWithModel is saved as upscaler and it requires upscaler_model.
sadly in future im probably going to caveat that I can't help debug custom nodes, people are going to have to go to the appropriate github pages with those issues.
Great tuto, very useful, thanks..
please also make one for using multiple loras and how to use the LoraStacker node.. Thank you :)
sure, ill take a look at lorastacker in the next couple of days, im fairly sure you can just connect lora in series by the way, just put them between the model and cliptext.
@@ferniclestix yeah but that's kind of annoying and confusing when working with large workflows that build pretty quickly.. so I'm curious about LoraStacker
th-cam.com/video/WFgf1I1aBxs/w-d-xo.html here, i had a look XD
Another good one!
Embedding could be very interesting to know how to make work.
In general, as a beginner I would like to see a quick guide of what is the difference between all the "models" in the comfyUI folder, hypernetworks, lora, embedding, lycoris, etc etc...
thanks
mmm, wouldn't be all that useful to most people that one though. Loaders automatically point to the relevant folder.
@@ferniclestix well thanks anyway :D
great tutorial!
Is there a good solution for the resolution problem with Reactor/roof and SDXL?
Guess the model those solutions use is just too lowres, so you don't get good results with the new higher resolutions SDXL provides. Faces always have less resolution than the rest of the image.
Also the face mask seems to include part of the hair most of the time, which also gets lowres by the process.
You can add additional steps with base or refiner afterwards, but if you use enough steps to fix the low resolution, the effect of roof is almost gone.
Are there similar models with higher resolutions from other developers or could we implement the model with Facedetailer (instead of a text input), which at least allows more accurate image masks.
Otherwise a good overview - found the same solutions when doing my research.
Good one, but the Reactor face swap won't open for me. There seems to be a bug around that node.
ive had issues with it before as well, suggest heading to the reactor github page and checking if there is a specific bug there may be a solution, if not you can post there and they might be able to help you out.
unfortunately due to the variability in comfyui installs, I can't really help much if a custom node is breaking for you. just make sure your comfyUI is fully up to date and your custom node is also up to date, and post whatever errors it throws up on the relevant github page.
Thanks! 👍🏼
thanks so much !!!!!!!!!
Fantastic vid!!!
i've inatalled ReActor but it doens't show when i load up comfyui. Impact is fine though, but it's the face swap i needed. running it on colab
have you checked out the github page for reactor? they fixed alot of the issues they did have.
hey love your work! i run into a issue when i upload your json files. it says i do.t have reactor notes. but i have it installed in the manager.why?
Reactor can be dificult to install, go to thier github page (you can click on the node name in manager to go there) there are some install instructions that may help if you are having issues. also obviously make sure you have restarted your server and browser after installing any custom nodes.
im going to try it again now
@@ferniclestix
The new version of Facedetailed is different from the one used in the video, there are 3 more options, "Guide_size_for", "noise_mask" and "force_inpaint". it blocks the rendering, because they are not connected and I don't know how to do it.
huh, ill look into this for you.
these are supposed to be widgets not inputs. right click on your node and switch them back to widgets, maybe they got accidently changed by accident.
Thanks vor your work. Sadly i have big troubles with the node. install seems to be brocken. i have to reinstall comfy:(
ahh that sucks, yeah, ive had to reinstall comfy... twice now? while installing custom nodes.
Sadly, I cant really teach how to python because im absolutely crap at it lol. I've fumbled my way through some fixes though, mostly its a result of nodes installing the wrong version of modules, conflicting with eachother or just plain installing to the wrong place. sigh, its so fiddly.
Couldn't get reactor face swap to install correctly for ComfyUI. When I run its install.bat, it whines about Cython not being installed. Open source software really can be a mess sometimes lol
true, if you have major issues, its a good idea to post any errors on the nodes github bug report thing, best way to get fixes and responses. sadly one of the things I can't really help much with is when a custom node doesnt work... you know, because it works for me so without knowing the ins and outs of the node and python cant really delve into it to find problems.
@@ferniclestix thanks nonetheless
Amazing video! Could you make some tutorials about batch i2i for making animations from a real video?
animations are on the list, not for a while though, ive got to do a couple other tutorials first... probably prompts then compositing, then I might be up for either animaton or an indepth tutorial on impact pack.
Awesome video again, I'm getting a "Error occurred when executing MMDetDetectorProvider: No module named 'gradio'" error, just wondering if this is some easy setting I have wrong or something.
hmmm... looks like maybe impact pack has an install issue? mmdetectorprovider is an impact pack node. I've had plenty of errors but not this one before.
@@ferniclestix Thanks but it's something basic I think, it's happening when it tries to load any type of upscale model now. I 've had it before I think but...🤬
could be an install bug with gradio too, its a module in python so... maybe you have system path issues or something. I had issues liek that aaages ago, ended up reinstalling windows to make it shut up XD
needs a section on procuring the nodes first though, some links would be even lovelier
mm, I struggle with installing nodes sometimes, I generally stay away from the topic of how to get them going because in all honesty, I succckk at python and getting it to do what I say. I don't feel confident giving advice about installing nodes because of this and its so tedious debugging them when they break. Ill always now just say use ComfyUI manager and visit the relevant github page if it messes up.
@@ferniclestix ohhh, I get you, no worries big homie💯
@@ferniclestix how do you do multiple control nets i need it for a project. plz help
efficiency nodes has a node called control net stacker that can make it easier but you can connect them in series although you may have to balance the strength to get it to apply sanely.
FaceRestoreWithModel always error.
many of the face restore nodes have issues. It's a result of using the portable install of comfyUI.
the standards for comfyUI node makers arent high and many forget to add the possibility of embedded python or system python being options for comfyUI and sometimes hardcode installs to the wrong place. or when trying to install we mess it up by installing modules to the wrong place.
reactor is not coming help please
I'd need more of an explanation to help out
thank you for this video ,i have one request , is it possible to make a video using cut-off region node by blender neko
Ill look into it, got a few things already lined up tutorial wise though.
no problem ☺@@ferniclestix
th-cam.com/video/ZZ1bc5T094s/w-d-xo.html there, had a look, make sure your sampler is set to normal scheduler :P
@@ferniclestix thank you so much for this.
I was really struggling with this node.
I didnt think the scheduler would be the culprit here😄.
I tried to make it very complex, say for example, i tried to make an image having a family of 4 with different dressing styles with different colors, no matter what i did the colors would bleed into each other horribly.
I am going to try it again, but this time change the scheduler.
Thank you once again for this😊
Thanks again
Np let me know if it doent stick, ill take a deeper look.
Im having trouble finding the VAE. Can anyone provide a link? :)
And also retinaface_resnet50 :(
I got it off hugging face but as its a non-standard version it may have been overwritten, there are a few different versions, the one im using is 'supposed' to reduce some artifacting, the vae that comes embedded with SDXL base should generally be fine.
@@ferniclestix Thank you. I got that part working I am happy to report. Any plans to maybe make a video about how to faceswap with roop?
technically reactor is a 'roop like' in that it is based on roop. I've tried to get roop nodes working but most are pretty busted.
@@ferniclestix Yeah, Ive had problems with roop. It wont even show up on Webui, just errors. I have been playing around with softwares like Deepfacelab and Faceswap but that stuff takes alot of time. Would it be possible for you to do a video on faceswapping in Stable Diffusion?
hi, great videos so far. I'm having trouble installing ReActorfaceswap, something during installation that wheel could not be build, Cython module is missing?
yeah, hang on, i got you fam, github.com/Gourieff/comfyui-reactor-node - down the bottom under troubleshooting, it tells you what to do. although the page is terrible at telling you.
Visual Studio or VS C++ Build Tools needs to be installed. There is a prebuilt option although I'm not sure how to install it to comfyUI, so its better jsut to install visual studio and make sure you install the python and C++ module
@@ferniclestix hmm, still doesn't work. Problem appears to be Insightface (ModuleNotFoundError: No module named 'insightface'). Perhaps I'll try another method for faceswap.
trust me, this is the fix, you need to install visual studio, I know this because I had the exact same error and thats what fixed it for me. ( think I also reinstalled the reactor node after I installed visual studio.
@@ferniclestix nope, did do the install end reinstalled reactor node, still an insightface error, and the reactor node won't start. Additionally it says: Error: OpenAI API key is invalid OpenAI features wont work for you. Is this related. Sorry, I am not very good at this Pythonstuff.
Im afraid I can't really walk you through a fix, I can only tell you what fixed it for me :( I think I had to do a module install using pip as well, I went to the node that reactor is based on and followed parts of those install instructions.
Open AI key bug is from having the quality of life nodes and won't affect anything. Its to do with some of the prompt AI tools it uses (gpt and stuff) it needs the key so it can use gpt for prompting... mostly useless tbh, I ignore it.
This is where I found the stuff that fixed my Reactor github.com/Gourieff/sd-webui-reactor - its part of the non comfyUI one, but the person who made the comfyUI version didn't bother doing a proper install instructions.
Stargate!
XD watching the series in chronological order.
when i try installing the thing, i have this error : module 'cv2.gapi.wip.draw' has no attribute 'Text'
you'll have to be more specific about what thing you refer to, haven't had this error myself, although error's in modules are usually due to an installation issue like not following install instructions correctly.
Great tutorial. Where did you get the face restore models like GFPAN or 4x-ultrasharp. Please share a link. Thanks
ahh huggingface.co and you gotta kinda hunt for them. theres also civitai I think I got mine through the EasyDiffusion installer which lets you ez install it in settings somewhere... like aaaaages ago. 4x ultrasharp was definately hugging face, and it can be a real pain to find.
Sorry I don't have links handy because my memory is that of a gold fish