Yolo segmentation seems to only detect one face

[Michoko92]

Hi,

First, thank you for this awesome UI! I'm trying to generate an image, where all faces will be fixed, like in the ADetailer extension of A1111.

Here is my prompt:

oil painting of a male actor in a theatre full of people in the 1920s, black short hair with gomina, wearing a hamlet costume, fantasy art by greg olsen
<segment:yolo-face_yolov8m-seg_60.pt>

However, it just fixes the face of my main character, not all faces in the crowd. I tried to lower the detection threshold, but still, only one face seems to be detected (other faces are clearly identifiable though, and ADetailer in A1111 works as expected).

oil painting of a male actor in a theatre full of people in the 1920s, black short hair with gomina, wearing a hamlet costume, fantasy art by greg olsen
<segment:yolo-face_yolov8m-seg_60.pt,0.6,0.2>

Do you have a suggestion to fix this, and have all faces detected, please? Thank you!

[FurkanGozukara]

I think we can have options like, top biggest 1, 2, 3 or like all detections

[Michoko92]

I think we can have options like, top biggest 1, 2, 3 or like all detections

Normally, if you use "-1" at the end of the model name, it will pick the first face identified. -2 means the second, and so on. If we put nothing, like I did, it should fix all detected faces, if I'm not mistaken. However, I like your idea, like using "+x" instead on "-x" to pick the x first faces that were identified.

[mcmonkey4eva]

You can test mask results with Advanced -> Regional Prompting -> Save Segment Mask

I tested that with prompt a photo of two women smiling <segment:yolo-face_yolov8m-seg_60.pt> beautiful smiling faces
and it very definitely masked both faces:

With your prompt, it skipped some of the background faces but got a bunch:

[Michoko92]

Hi,

Thank you for your reply and explanations. Saving the mask allowed me to play with this feature, and in some cases, I indeed have a few more faces detected. But nothing similar to ADetailer, which detects many more faces. In SwarmUI, only the most obvious faces, looking at the camera, are detected. Even if I try to lower the detection threshold to the maximum, it detects many fewer faces than ADetailer (only 3 or 4, while ADetailer detects dozens). Do you know if there is a way to make the Yolo model more sensitive? (as it is mainly useful for fixing smaller faces that are generally more deformed and less detailed). Thank you!

[mcmonkey4eva]

The only reason it'd be different in adetailer vs swarm when using the exact same yolo model, is if the threshold minimum of 0.01 is somehow too high (unlikely?).
More likely reason is just maybe your adetailer is using a different yolo model that you need to copy over

[FurkanGozukara]

Can we inpaint only largest detected face?

segment:yolo-face_yolov9c.pt-2,0.6,0.5

how to modify this?

[mcmonkey4eva]

Open a separate feature request for that