This model is designed to detect pairs of eyes, unlike this model https://civarchive.com/models/150925/eyes-detection-adetailer , mine is designed to detect pairs of eyes. Thanks to the creator, this has inspired me to train this model
This is my first project on yolov8
Recommended confidence threshold 0.3-0.4
The model works:
1) With anime characters;
2) With humans;
3) With animals;
4) with anyone with pronounced eyes;
5) Works at close and very far distances.
The model also detects closed eyes or if one eye is open but both eyes are visible.
What doesn't work:
1) If a character has one eye covered by something (patch, hair, etc), the model does not recognize it;
2) If there is only one eye in the image with no context that there is another eye.
How to use:
How to install:
Go to https://github.com/Bing-su/adetailer and install Adetailer
Place the this file (full_eyes_detect_yolo8s.pt) in "stable-diffusion-webui\models\adetailer"
and reload WebUI.
Install for Automatic1111:
Open "Extensions" tab.
Open "Install from URL" tab in the tab.
Enter https://github.com/Bing-su/adetailer.git to "URL for extension's git repository".
Press "Install" button.
Wait 5 seconds, and you will see the message "Installed into stable-diffusion-webui\extensions\adetailer. Use Installed tab to restart".
Go to "Installed" tab, click "Check for updates", and then click "Apply and restart UI". (The next time you can also use this method to update extensions.)
Completely restart A1111 webui including your terminal. (If you do not know what is a "terminal", you can reboot your computer: turn your computer off and turn it on again.)
Description
Did a massive job, manually marking up 2k images. 1k realistic and 1k anime pictures
FAQ
Comments (11)
how to use in comfyui ?
after using this what we get? means distorted eye correction?
The model is used for additional processing of the eyes, it is possible that one face processing is not enough and it is necessary to go over the eyes additionally. This is why the model exists.
About installation, I don't use Comfy myself, but I will give you a link to a comment where a man tells how to do it with a similar model.
https://civitai.com/models/150925?modelVersionId=168820&dialog=commentThread&commentId=318992
This is awesome! Works perfectly.
I need to ask, I dont know anything about training, would you mind trying to do a "smiling" one? or just mouth I dont know, the idea being of detecting the mouth to fix crooked smiles/teeth and such.
Thanks again for your work
How is this different from using the old 'individual' eye detection with a 'combined: true' in the Mask to SEGS node?
Well I guess it would work better for pictures with multiple people, but that's the only situation I can think of?
Does/Can it default to one eye only in case where the other is obstructed and/or out of sight?
I'm wondering this myself. Perhaps detecting them together promotes the generation to make them more similar, since they are detected as a pair belonging together? I'll give this a whirl when I get home from work, compare results.
Any chance of Safetensor format?
not necessary. especially if using Automatic1111 which has a safe unpickle function that can't be turned off unless you specifically write it into startup arguments.
Turns out I've ended up using this ADetailer model A LOT because other eye models will take each eye individually resulting in differences. This is my 3rd most used ADetailer model after Bingsu's "person" and "face" models - and that's saying a lot. Great upload, thank you.
This is amazing, and I use it it allomst all generations, however, it has a tendency to change the skin around the eyes and basically everywhere within the detection rectangle. I've tried every way to limit this but it is stubborn. In half of those images though, the change to skin is subtle
old comment, but you can resolve this with one of these changes, or some combination of them (given in order of what I think is the best setting to change first):
- significantly raise your "Inpaint only masked padding", which will give the denoiser more context to match accurately
- raise mask blur, which will specifically blur the change in skin making the change blend better
- lower denoising, preventing the denoiser from having the latent space to alter the skin enough
- lower mask dilation value, which is by itself kind of like making all three of the above changes at the same time, but in a more dangerous way
top