@YukinoAI I am making a clothes-replacing app and needed an eye detection model to inpaint a superhero mask on top of the person's eyes. Now that I have tested your model, it only seems to work well for anime pics. Still thanks

Not a issue. Just set it to 'merge' mode and they will be inpainted at the same time.

@YukinoAI Consider making a feature request on the Adetailer github to sort detections by the X axis. I believe there is already a way to split the prompt across each inpaint though I've never tried it: link - so we only need to be able to sort them.

@Jilhbril Not sure how helpful this is but in case of regular ADetailer, the use of [SEP] seperates out characters when prompting multiple people, maybe it would work at least on a single person prompt in the same way?

@Jilhbril yep i found that out shortly after making this post :)

ADetailer, once installed and enabled in your UI, will be a drop-down section below your image generation parameters in A1111. Expand it to find Enable ADetailer at the top, and settings below. These will be various options to select portions of an image to run a quick inpainting sequence on at the end of your generations when enabled.

@noahspeaks257 Ok thanks for the info! But I don't see how this is different than just masking the exact area you want for inpainting. Is this for people who can't draw with a mouse or something?

@pkayze Its for people who dont want to do it manually hundreds/thousands of times. By that kind of logic why use SD when you can just draw it yourself?

@Jilhbril That's nowhere close to the same analogy. I don't see any batch function in the videos. It seems this is just picking the area for you (with varying confidence, meaning sometimes it may not do it properly) vs. you just putting clicking once with the mask button over each eye. Admittedly I haven't installed it and maybe there's more to it, but unless there's a batch feature where you can set it to it automatically and run on a batch of pics, I don't see how this is better.

@pkayze Do you think you have to adjust its settings every time you change the prompt or something? The 'batch feature' is setting it up right and leaving it enabled while you keep making images and it will inpaint them for you automatically. You can adjust the confidence threshold so only what is very likely to be an eye will be detected and you can also limit it to the top 2 detections so it will ignore characters in the background etc... This is a very powerful tool that even has API and ControlNet support. Its pretty much automated inpainting without the manual labor if that's not something you are interested then that's fine I guess.

EDIT: if you want to use this on existing imgs you would need to use the API since indeed a 'batch' feature like we have in the 'Extras' tab does not exist yet. Something to consider for a PR.

Big thanks!

Thanks, I can't figure out how to get it to do both eyes. It only does 1.

EDIT:

Sorry, I overcomplicated things by also using the samdetector. It works wonderfully!

@haraririri So to be clear do not use samdetector just bbox?

They process in order, so do the face first and then the eyes

I thought you were supposed to put the lora reference in the detailer's prompt, under the adetailer tab. that seems to work for me. don't ask me though.. I'm a n00b

@YukinoAI Thanks! and I can't wait for the hand detection!

I guess itis the part of "Mask only the top k largest (0 to disable)" because that's the option to control how much eyes you want to be detected.

Try to add colors you don't wanna see in negative in adetailer or lower denoising strength a bit,

@Marader Problem is that it's the same color just vastly different shades.
I'll try lowering the denoising though.

@SirDigsbey There's an option in settings, image.png (917×95) (discordapp.net) add the "img2img_color_correction" and if you see this checkbox active image.png (513×64) (discordapp.net) try to disable it and see the results, i think i have similar problem and this relove it. Well that's for a1111, if you using another UI there's probably similar option.

@SirDigsbey https://civitai.com/images/3494299 I have almost zero issues with that workflow. The trick is among other things to set the mask to merge, otherwise it paints both eyes without any context for the other.

@Sophorium i think the problem was not with the workflow, but with keeping original color settings.

@Marader Ok well if it's not heterochromia then I dont know, other than reinforcing eyecolor on the adetailer prompt itself, mine are usually the same shade

In the ADetailer tab, there is a Mask Preprocessing tab, right?

Why don't you set the Mask merge mode to merge in that tab?

I'm using translation software, so sorry if the sentences are strange.

@yamage77755 didn't know you can merge two mask that not connected😲

Specify an eye color in a custom prompt for AD, without a bunch of (positive) chaff. Without a custom prompt, it's trying to put the contents of your entire prompt in every eye, and the color you get is whatever depending on prompt, denoising, etc. (Don't recommend merge, because then you change the direction the eyes are looking depending on padding->mask dilation and camera angle.)

@SnowyYukino Compared to any face_yolov8 model is doesn't make changes... Even increasing denoising strenght with main values is not working for me, I don't see real changes. That's why I asked if I had to do something apart from installing. I basically wanted a model like face_yolov8 but only for eyes. Those work the best, but sometimes I don't want to change the expression. Probably yours makes very little changes?

@SnowyYukino Okay... This is what I meant: Here I tested all similar models to your pt model. Imgur: The magic of the Internet
You need to select MERGED otherwise it won't show the best performance. I've seen another guy posting a picture with the recommended settings. I think you must say it in the description of your model.
And even then, it is not better than Adetailer in my opinion... Probably because Adetailer has a better inner model for eyes.

use face detailer (pipe) that comes with comfy impact pack, then move this model in comfyui models in bbox found in ultralytics (so basically move this model in /models/ultralytics/bbox/) then use it as if u are using face detailer (if u don't know how to use it, check facedetailer in comfy impact pack in github, they have a detailed guide with workflows) and it should work like normal face detailer, except now it uses eyes models, so it will detail eyes, i tried it and it's working perfectly, good luck!

Awesome, sauce! Thanks a million that makes a heap a sense now that I think about it lol,

@boudy In Arnold's voice from Predator. "Yooz SUN OF A BITCH!"

@boudy this create 2 differents masks and is treated 1 at the time in the detailer leading very often to 2 different eyes as for result. in adetailer in a1111 they are corrected simultaneously that lead to a more consistent result.

@bissonfrederic69429 Noticed. Did you find a solution?

@jahagafalakajaha656 with flux it's not a prob anymore, but back in the time, yes i was using a crop portion of the face to do the detailer then merged back only the eyes to the original image with some special node, mask composite or something

Click "Show More" in description, it gives detailed guide on how to install Adetailer extension and use the model for it

The eye detailer just detects eyes (including lashes, and potentially brows too if you expand it). Then when it inpaints, it will improve all of them.

This is segmentation\bbox model they are universal. From what I know civitai require to specify base model that's why is SD1.5 but this will work with any model since it doesn't affect generation (it's for detection on already existing image)

Can you provide the workflow please, i am having hard time following this.

@mynameisjimcarrey204 sorry, didn't checkup on this site for a while... I uploaded a rough demonstration workflow here

https://civitai.com/models/2317764?modelVersionId=2607550