Core Concepts
DiffuMatting proposes a novel approach to generate high-accuracy matting-level annotations for arbitrary objects, facilitating downstream tasks and enabling community-friendly art design.
Stats
DiffuMatting synthesizes general object and portrait matting sets, reducing relative MSE error by 15.4% and 11.4%, respectively.
Green100K dataset includes 100,000 high-resolution images paired with high-accurate annotations.
Quotes
"Our DiffuMatting serves as a highly accurate 'anything matting' factory."
"DiffuMatting combines the robust generative capabilities of diffusion and the added functionality of 'matting anything'."