The detected objects and the segmented objects are inconsistent

[magicwang1111]

[kijai]

So the issue here is that the SAM2 video mode only supports point prompts and not full boxes, the Florence2 Coordinates -node simply selects center of the box and passes that as a point to SAM2, in this case it would be the shirt and it segments only that.

The video mode also supports multiple points, so in this case the Florence2 box detection simply won't work well.

[tristan22mc]

Is it possible to give some kind of a prompt to SAM2 so it has a better idea what to segment? I was having an issue segmenting a product out of an image and it kept selecting the text instead of the whole product

[kijai]

Is it possible to give some kind of a prompt to SAM2 so it has a better idea what to segment? I was having an issue segmenting a product out of an image and it kept selecting the text instead of the whole product

Prompting for SAM is just points, boxes or masks, it doesn't have any innate way to handle text. It works best with the points, for which I have been working on the PointsEditor -node now included in the video example too.