What's the different to the "segment anything + inpainting" pipeline?

[longcw]

Thanks for sharing this cool demo. It looks similar to the pipeline of masking out the foreground and then inpainting the background, optionally with some IP adaptor-like background control. Since there is no paper released, what's the difference between that? Is it possible to maintain a consistent ID while also allowing for deformation such as automatic adjustments to shape, size, and viewing angle?

[chenbinghui1]

@longcw
Thanks for your attention. SAM and inpainting are basic tricks in our pipeline, while we propose many other improvements for the generation of background/human/human-object-interaction and so on.

Exactly maintaining object's ID and automatic adjustments of shape, size, viewing angle are trade-off theoretically. Now ReplaceAnything v1.0 cannot automatic adjust like shape, size and viewing angle. We are developing V2.0 for changing shape, size and viewing angle.

[longcw]

Thank you for the quick reply. Indeed making the inpainting stable with high quality is nontrivial already. Looking forward to the paper and the v2, also hopefully for the code :)

[GallonDeng]

also looking forward to the paper and the v2, v1 too slow and the size not right