Put the model under models/adetailer
Only tested with realistic and semi-realistic images. There is no anime / cartoon images in the training data.
Description
This version is not meant as a replacement for 1.0 but is just overall more sensitive and will more likely detect obstructed or deformed breast. It'll also have more false positives, so only use this if you feel like version 1.0 doesn't detect the breast.
FAQ
Comments (25)
perhaps an off topic question but has anyone managed to force adetailer to run last of active extensions?
I've tried renaming the folder but thaft only affects ui
Renaming the folder should work, but I think certain extensions don't run in a specific order like Controlnet.
using img2img with adetailer instead of txt2img worked for me, also the same with inpainting. Only problem is that sometimes Adetailer generates an entire body in the inpaint zone hahah
When i try to run the 1.1 sentitive model, i get this error:
WARNING models\adetailer\female_breast_v1.1-sensitive.pt appears to require 'ultralytics.nn.modules.conv', which is not in ultralytics requirements.
AutoInstall will run now for 'ultralytics.nn.modules.conv' but this feature will be removed in the future.
Recommend fixes are to train a new model using the latest 'ultralytics' package or to run a command with an official YOLOv8 model, i.e. 'yolo predict model=yolov8n.pt'
requirements: YOLOv8 requirement "ultralytics.nn.modules.conv" not found, attempting AutoUpdate...
Try this: https://github.com/ultralytics/ultralytics/issues/2573
Otherwise:
On which ADetailer version are you?
Does the 1.0 version work? Do the default ADetailer models work?
pip3 install --upgrade ultralytics solved this
@Jellon the face eyes and person models work fine.
@sci Maybe they trained it on an older version of ultralytics. I'm using 8.0.119, so want to have a newer version than that. Check out the link I've provided and the solution NiKiuS posted.
doesn't seem to work. Detected 2 out of 10 with threshold 0.1
Can you provide an example of what it didn't detect?
@Jellon https://imgur.com/a/RxOwOA2
NSFW. First one it only detect 2 pairs. Second one, detect 0
@Wideanon The Problem here is that ADetailer is using the default image detection size of 640, which is simply too small for your image size.
The default face model also didn't detect anything in my test:
0: 320x640 (no detections), 93.3ms
Speed: 2.0ms preprocess, 93.3ms inference, 1.0ms postprocess per image at shape (1, 3, 320, 640)
[-] ADetailer: nothing detected on image 1 with 1st settings.
0: 320x640 (no detections), 12.0ms
Speed: 2.0ms preprocess, 12.0ms inference, 1.0ms postprocess per image at shape (1, 3, 320, 640)
[-] ADetailer: nothing detected on image 1 with 2nd settings.
If I run a manual test with image detection size of 2560 here's the result:
@Jellon figured so. I used comfyUI facedetailer to test, pretty much the same thing as adetailer. It seems you can't change detection resolution either.
@Wideanon You could change it manually in the code.
It's in extensions/adetailer/adetailer/ultralytics.py line 22:
pred = model(image, conf=confidence, device=device, imgsz=2560)
But if you do that don't forget to remove it afterwards. imgsz of 2560 won't work for small images.
@Jellon What about when they USED to detect, then they all of a sudden stopped on the same image? 832x1040 is still too high?
@Kaladae Are we talking about the same model that used to detect, but doesn't anymore? If that's the case I'd assume a setting in ADetailer is wrong.
But I'd need to see the image and run it through the manual detection to debug.
@Jellon It's actually for an eyes model, I've been running through all the adetailer comments to try and find clues. I got it set up in AUTO1111 and made a few generations, was working great, fine tuned prompts, deleted a couple pickle-positives (that may be what did it) and updated adetailer and grabbed a few pickle-negative tensors and now I have to set the detection from 0.3 to 0.01-0.03 to get a single eye (out of 2, ofc) to be recognized, some of the time.
So it's not this one, I'm just having difficulty overall. Your comment was the first to lend me a little insight into the possible issue, but I did lower the resolution and it still kept failing.
@Kaladae Eyes are probably very difficult for everything that's not a portrait. Generally anything below 0.3 is not considered a detection. I would never lower the threshold below 0.3. It sounds like something is wrong with the model, but it's really hard to give advice without some examples. I'd contact the model maker and provide some example so they know what you're trying to do.
@Jellon This was a portrait style, and like I said it worked fine the first few uses. Then I tweaked some things, and it stopped, so I'm guessing I messed something up. Thanks for giving me some leads though. edit It randomly started working again after changing the particular model. I guess it's related directly to one of the models I was using...
Is there any way to force it to only detect breast pairs and not single breasts?
I've been having issues on random pictures of it detecting single breasts and then generating 2 breasts on that single breast :/
No unfortunatelly, this is something ADetailer doesn't support (yet). If it generates 2 breasts I think your denoise strength is too high or your prompt is not well designed.
@Jellon What kind of prompt would you recommened? Currently my prompt is simply: "natural breasts, small nipples" for Adetailer
@pimmel5 Remove the nipples part, I think that causes issues. And make sure your denoising strength is below 0.4. I'd experiment with normal inpainting first, if your prompt works there it also works with ADetailer.
Change the mode to "merge" for "Mask merge mode" in Mask Preprocessing if you're using Adetailer.
@32Bitshifter Oh, how does that affect the mask preprocess?