J on Nostr: nprofile1q…5g9hk So here I have put together segmentation with Sam2, I find it a ...

nprofile1qy2hwumn8ghj7un9d3shjtnddaehgu3wwp6kyqpqu0u38hc0gdpkmjdyjv3kyd928x4jp0s4xjc9d2lwjjxawrytn7fq45g9hk (nprofile…g9hk) So here I have put together segmentation with Sam2, I find it a lot better than running each frame through segm detector as this will follow the subject a lot more closely. It kinda goes back to what I said first about just cutting out the character and running that through your animdiff pipeline, and then compositing her back in afterwards (you are using depth maps so the shape of the character should be mostly the same), will send this workflow shortly.