More training on "people/animals" for the Magic Mask

I imagine that BMD is continually refining their object recognition models for the MagicMask tool.
Super smart people are working on these things, so what I'm writing is just general ideas thrown out there.

It is my assumption that these models are trained on "objects" appearing on screen, rather than specific ones.

When the Magic Mask tool performs well, it does so impressively. However, I believe that the transition from subpar performance to achieving a flawless selection should be more seamless.

I've mentioned this concern previously, but I sincerely believe that labeling this tool as "better" can be misleading. In my experience, the "better" option often falls short compared to the normal mode, especially with some adjustments like slight edge blurring, etc. With "better", more background elements are picked up.

I propose the introduction of a dedicated object selection mode designed exclusively for humans and animals. This mode should be trained exclusively on these two categories, and other elements should be excluded. The models should be trained on video clips with low contrast, minimal distinction between the background and foreground, without any depth of field (such as a head with brown hair passing in front of dark elements in the background, things like that).

Alternatively, these improvements could be applied to the "person" selection mode (because it's still the old one and it's way lessprecise for edges). This mode is not really for rotoscoping anyway, but for a general selection of someone's body part. It could benefit from an update using the same technology as the object selection mode, but with a focus on humans (and possibly animals) only.

Example of segmentation used to direct image generation like Stable Diffusion

0_v7vQw0H_QnqjBX4p.jpg (52.58 KiB) Viewed 408 times

And for the love of God, please allow use to upload .webp images [The extension webp is not allowed.] - It would take less space on your server too