The goal of this point is to find a face of any type of generation, including realistic, anime, and anthro characters.
PLEASE NOTE THE DATABASE TYPE DOES NOT MATTER. I have adjusted each version to include a different database version for search ability, but these are diffusion agnostic. It's all about detecting a face no matter what.
This is a bbox detection YOLOv8 model. This does not improve your image on its own and must be used as a part of the 'adetailer' extension for WebUI, found at https://github.com/Bing-su/adetailer or 'uddetailer' found at https://github.com/wkpark/uddetailer. It is to be used to find any type of face in a image so that you can infill that part of the image. Right now it is better suited for finding faces that the other pre-installed databases seem to miss. (ex. face_yolov8n.pt, face_yolov8s.pt, mediapipe_face_full). If those did not work for you, mine probably will.
This is my first model and first item I have made I feel confident in that others may find it useful. I found out that these adetailer databases are actually part of something that I was already working on.
Right now it is not trained on realistic animals or common mistakes, but I have seen my desired success rate already. Please let me know if you continually get a false positive (what you and I know probably should not be a face).
I realized to late that I focused to hard on if I could, that I didn't stop to think if I should. Now I am making a new data set and I have to re-tag all my 'facial' data... The next one will have classes. There are some rumblings that adetailer may be able to pick and choose which classes of detection to remake.
Later if I can work this out I'll probably move on to segmentation, time willing. I have met challenges with this with my tools.
Please feel free to provide your feedback on which faces it seems to miss or what false positive you find, as I want to continue to improve this model to find Only faces :)
Description
V1.2 has a significantly reduced data set. This comes with the benefit of significantly less false positives, and it can find a variety of faces among humans, anime, and animal. However, the data set did not include data of 'bad faces', so if the face is to weird or to small, it likely will not find it as I found in my testing.
FAQ
Comments (49)
can you make a hands finder
There's a few already, but what are you hoping for? Realistic human, cartoon/anime, animal paws, dragon talons?
I think I'll take a stab after I finish the rework of this one, which unfortunately takes a while as it is just me. I find it oddly fun but it's still very time consuming. On the off-hand (pun intended), if you know of a large repository of generated images with hands good and bad, lemme know as that is half the trouble, finding images to train it on the concept.
Also teaching computers about hands is a difficult concept already, which is why it messes them up so much. In image recognition like this, now we aren't just looking for good hands, but also possible mess-ups.
@l0ud_ninja oh ok thank for the respond
@l0ud_ninja A paw model would be nice, probably easier then hands too.
@xehava3 what paws would you be interested in? dog, cat, bear, etc?
@l0ud_ninja Canine paws.
I seem to be confused with the installation. Does this go into the adetailer models folder? I placed it in there, and it doesn't appear as an option in the dropdown list.
Yes, it goes in webui/models/adetailer. Then you have to restart WebUI for it to show up.
I would really love to see a version specifically for realistic animals (especially canines) as around half to two thirds of the images I generate don't detect the face.
I observed that using Adetailer with SDXL models (both Turbo and Non-Turbo variants) leads to an overly smooth skin texture in upscaled faces, devoid of the natural imperfections and pores.
On the other hand, the combination of Adetailer with SD1.5 models noticeably enhances the skin's details and imperfections.
This conclusion was drawn from extensive testing with multiple SDXL Turbo, SDXL Non-Turbo, and SD1.5 models. Despite experimenting with various Adetailer settings – including adjustments to resolution, steps, denoising, and trying different prompt approaches (full, none, pruned) – the issue with SDXL remained unresolved.
Below are some side-by-side comparisons (the images on the left are without Adetailer, and those on the right show the results with Adetailer )
RealvisXL 3 Turbo:
https://i.imgur.com/ET1JXpQ.jpg
https://i.imgur.com/PisIdpA.jpg
RealvisXL 3 Non-Turbo
https://i.imgur.com/tyZZdgI.jpg
https://i.imgur.com/j8uco63.jpg
SD1.5 Cyberrealistic v41 and AnalogMadness 7
https://i.imgur.com/xwhBL8i.jpg
https://i.imgur.com/xWBKp5m.jpg
Adetailer settings: default
https://i.imgur.com/5ko6BX8.png
Why is there such a difference? Is Adetailer not compatible with SDXL? Any insights or shared experiences in this matter would be greatly appreciated.
Adetailer is definately compatible with SDXL. With that said, changing models to achieve better results it's very common practice and even encouraged! That's your artistic vision at play and you should strive to achieve unique and great results. This is the new revolution of Art, that's how you make the product perfect.
what settings do you use for adetailer?
@es69 I make sure to describe the thing I am trying to detect in the positive and clean it up in the negative, but keep it short. If the checkpoint is never good at faces, I switch to a better, but stylistically similar checkpoint. For example, If the first checkpoint can only do realistic, don't use one in adetailer that can only do anime. I always use DDIM and Uniform, as that seems to style match the best, while getting great levels of detail. Usually, if it couldn't be detected, it probably couldn't be much improved, but again that all depends on the checkpoint's ability. I hope those tips help.
AMAZING! You rock! I was struggling with furry characters from the pony model, but this helps so much! If you ever have more time, a paws/claws model for hands and feet would be fantastic! The normal hands one isn't that great for furry characters.... Anyhow outstanding work, thanks again!
Thank you! I'll go looking for some 'paws, claws, and automobiles' :)
@Toad_boi hey hey there, any progress kind sir? :)
Hey, I've been using your model for a while now and it's been working really well until recently. I'm not sure if something changed with Adetailer or the latest update of Automatic1111, but when I try to use your model with Adetailer, it works for a few images than causes a hard crash. The crash only happens when Adetailer tries to detect faces after an image is generated. This happens regardless of checkpoint, vae, loras, sampler used or which extensions are enabled. I've tried several other Yolov8 models with Adetailer and this is the only model that causes this. I'm not really sure what I need to do on my end to fix it, and I've been avoiding since I found out. I'd appreciate any help with this, and I can send my terminal output if you need it.
I'll look into it! gimme your dump. However, I have also had a similar issue as you describe on other yolo models.
@l0ud_ninja Here's my dump. Glad to hear I'm not the only one having this problem. I put my hardware info at the top in case you need it. Thanks for taking a look!
♥ ♥ ♥ Thank you so much! ♥ ♥ ♥
i'm having fun using this and it's been great so far. except i noticed that it's always failed to recognize the face if the character was looking behind. are there any chance to fix this issue or do you have tips for this one ?
oh, you mean like looking over the shoulder? most of the training set was where the character is mostly facing the camera. I'm looking for training data sets for more items, I'll see if I can find a good set for partial faces.
However, I do have an issue as well already with a recognition of chests/knees/elbows as faces?
@Toad_boi ah yeah looking over the shoulder, i can't remember the name for it 🤣.
Also yeah i do get some of that knee/stomach/chest recognised as faces too but it's not often
Its better for furry faces and seems to do the same job as the default AD, which is great.
It detects anthro faces almost always, even if complicated, BUT ferals is often very difficult. Sometimes it gets one instantly at like 90% certainty, sometimes it would rather detect a random part of a tree as a face then a dragon, or anything else with a long snout, even when looking forward. When i put detection to 0.25 or lower, it can mark like 5 different things, still none of which are the dragons face.
To prove that, heres a screen of it detecting 3 things as faces... ON a dragons face. But not the face itself. On 0.2-0.1 value! It should detect like the tiniest smidge of a face at that low confidence. Which it does. Just not an actual face.
Which version did you use? Different versions used different data sets and training methods. Try using one of the other versions if you're attempting to detect more dragonic / horse / symbiote / horror faces.
I'll make sure to include more facial variety on another test run
@Toad_boi Wait other versions might be better? I thought because they went higher in number they were just upgrades, and used 1.2. I guess i can try them, but i also didnt wanna bash you, it prob sounded that way because i was just so done at it not detecting my dragons faces no matter what. Like, the picture i posted was basically mogging me lol.
My guess it that it simply lacked faces with long snouts and reptilians, like dragons, lizards, crocodiles to train on. I might be wrong, but that would just be the first thing i would try that seems like it could be an easy fix. c:
This solve my problems detecting pony faces, love u.
Love u, 2!
For some reason i can never get ANY Adetailrs to work (despite being in the correct folder, and actvating adetailer manually each time i load up forge) except the 'small face' one that comes default.
No matter what settings i tweak, or what i do.
How the hell do i get Adetailer to work? including this one. Hasn't worked once and it MUST be somethign i'm NOT doing or somehting..
ONLY the 'small face' one works and even then, not always.
Do any of the other files show up as options to select?
@Toad_boi I have all the defaults, + a handful of others in the dropdown menu.
It's not that theyre not showing up, just they don't actually trigger (Again, except the small face default)
I've messed with the detection settings on all of them setting confidence anywhere between 0.1 and 1, changed confidence to area, and vice versa.
@WorstAirtist what checkpoint are you using? Does the console log any error when you're generating images?
@Toad_boi To the second half of your comment, no. Literally just nothing (and to be clear this is on everything except the small face one that comes default and even that only works some of the time)
As for model... it depends. I use WaiNSFW as my primary, Steinillustrious, Dvine, and a few others.
@WorstAirtist sorry for the delay. I've looked around and I'm not sure. I suggest maybe delete your venv folder so that WebUI 'restarts', as that has fixed lots of issues for me before
@Toad_boi No need to apologize, it isn't a 'you' issue, just happened to be the one where the straw broke the back on my own lack of understanding.
I'll try it soon. Cause it makes no sense why my adetailer never works outside of a single version ~.~; will try it soon.
What WebUI are you using?
If you’re using reForge, for example, ADetailer won’t work without tweaking a couple of files due to "security risks."
Sometimes, ADetailer leaves a black box or lines around the affected area because of certain VAE arguments in some checkpoint models. Using --no-half and --no-half-vae can cause this issue.
Also, if you're reusing a picture/settings as a prompt and ControlNet is enabled, try disabling it and generating again with a new seed.
I hope this helps!
@Hyokkuda I am using normal Forge (most recent version as of 3 weeks ago)
I don't usually enable controlnet as I've never used it. and afaik its off by default both Cnet and adetailer cnet.
As for boxes.. I'm not even getting the boxes at all (except with mediapipe face short, which is the only one that works at all out of all 12 of the options i have) but it never leaves any 'traces' when its done.
I will try with a new seed as thats the only thing i can think of that might be the issue... potentially.
Regardless, thank you both for oyur answers.
Excellent model. Helps with all kinds of faces
isgood!
Hi, excuse me.
How Can I implement the detailer into comfyui?
This is good, but hope for a fix, since it detects dicks/testicles, instead of face sometimes im not sure why. . . it's funny though,
The infamous "dickhead".
@WetnWild im not sure if it's my own setting or what but it literally detects the shape of that whole pp 🤣
is it meant to say FacesV1.pt
Good at detecting faces, although depending on the art style and the inpaint masked padding it can mess the skin color. For example, with 32 padding pixels a character with a purple skin color the inpainted area results in a different shade of purple.
A style mismatch is an unfortunate consequence of inpainting, but I’m glad it’s working for you. Remember to fiddle with the confidence level on certain styles, lower confidence if it can’t find them but higher if it keeps finding things like armpits and knees as faces