adetailer 24.3.3

An object detection and auto-mask extension for stable diffusion webui.

ADetailer

ADetailer is a extension for stable diffusion webui, similar to Detection Detailer, except it uses ultralytics instead of the mmdet.

Install

(from Mikubill/sd-webui-controlnet)

1. Open "Extensions" tab.
2. Open "Install from URL" tab in the tab.
3. Enter https://github.com/Bing-su/adetailer.git to "URL for extension's git repository".
4. Press "Install" button.
5. Wait 5 seconds, and you will see the message "Installed into stable-diffusion-webui\extensions\adetailer. Use Installed tab to restart".
6. 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.)
7. 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.)

You can now install it directly from the Extensions tab.

You DON'T need to download any base model from huggingface.

Options

Model, Prompts
ADetailer model	Determine what to detect.	None = disable
ADetailer model classes	Comma separated class names to detect. only available when using YOLO World models	If blank, use default values. default = COCO 80 classes
ADetailer prompt, negative prompt	Prompts and negative prompts to apply	If left blank, it will use the same as the input.
Skip img2img	Skip img2img. In practice, this works by changing the step count of img2img to 1.	img2img only

Detection
Detection model confidence threshold	Only objects with a detection model confidence above this threshold are used for inpainting.
Mask min/max ratio	Only use masks whose area is between those ratios for the area of the entire image.
Mask only the top k largest	Only use the k objects with the largest area of the bbox.	0 to disable

If you want to exclude objects in the background, try setting the min ratio to around 0.01.

Mask Preprocessing
Mask x, y offset	Moves the mask horizontally and vertically by
Mask erosion (-) / dilation (+)	Enlarge or reduce the detected mask.	opencv example
Mask merge mode	None: Inpaint each mask Merge: Merge all masks and inpaint Merge and Invert: Merge all masks and Invert, then inpaint

Applied in this order: x, y offset → erosion/dilation → merge/invert.

Inpainting

Each option corresponds to a corresponding option on the inpaint tab. Therefore, please refer to the inpaint tab for usage details on how to use each option.

ControlNet Inpainting

You can use the ControlNet extension if you have ControlNet installed and ControlNet models.

Support inpaint, scribble, lineart, openpose, tile, depth controlnet models. Once you choose a model, the preprocessor is set automatically. It works separately from the model set by the Controlnet extension.

If you select Passthrough, the controlnet settings you set outside of ADetailer will be used.

Advanced Options

API request example: wiki/API

ui-config.json entries: wiki/ui-config.json

[SEP], [SKIP] tokens: wiki/Advanced

Media

• 🎥 どこよりも詳しいAfter Detailer (adetailer)の使い方① 【Stable Diffusion】
• 🎥 どこよりも詳しいAfter Detailer (adetailer)の使い方② 【Stable Diffusion】

Model

Model	Target	mAP 50	mAP 50-95
face_yolov8n.pt	2D / realistic face	0.660	0.366
face_yolov8s.pt	2D / realistic face	0.713	0.404
hand_yolov8n.pt	2D / realistic hand	0.767	0.505
person_yolov8n-seg.pt	2D / realistic person	0.782 (bbox) 0.761 (mask)	0.555 (bbox) 0.460 (mask)
person_yolov8s-seg.pt	2D / realistic person	0.824 (bbox) 0.809 (mask)	0.605 (bbox) 0.508 (mask)
mediapipe_face_full	realistic face	-	-
mediapipe_face_short	realistic face	-	-
mediapipe_face_mesh	realistic face	-	-

The YOLO models can be found on huggingface Bingsu/adetailer.

For a detailed description of the YOLO8 model, see: https://docs.ultralytics.com/models/yolov8/#overview

YOLO World model: https://docs.ultralytics.com/models/yolo-world/

Additional Model

Put your ultralytics yolo model in webui/models/adetailer. The model name should end with .pt.

It must be a bbox detection or segment model and use all label.

How it works

ADetailer works in three simple steps.

1. Create an image.
2. Detect object with a detection model and create a mask image.
3. Inpaint using the image from 1 and the mask from 2.

Development

ADetailer is developed and tested using the stable-diffusion 1.5 model, for the AUTOMATIC1111/stable-diffusion-webui repository only.

License

ADetailer is a derivative work that uses two AGPL-licensed works (stable-diffusion-webui, ultralytics) and is therefore distributed under the AGPL license.