Accuracy

[JohnSundayBigChin]

Hi, this is not really an issue, I just wanted to know why there are differences in the segmentation/masking/rotoscoping results between the published models run locally and the one obtained directly from the META page. Is it because of the model implementation, computing power or a Comfy UI limitation?

[kijai]

There shouldn't be any differences (unless there's been an update) in the detection and the raw mask, however this node doesn't do any postprocessing, so the raw output isn't comparable to the official demo. There are other nodes that can do post processing, such as filling holes or smoothing the edges, for example my GrowMaskWithBlur.

[JohnSundayBigChin]

Thanks for replying, I'm going to try that node now.
By the way, you've done an excellent job, much appreciated for sharing.