Główne pojęcia
Doob's h-transformを活用した新しい条件拡散プロトコルの提案と、その効果的な実装による画像生成とタンパク質設計での成果を示す。
Statystyki
dXt = bt(Xt) dt + σtdWt, XT ∼ PT,
dHt = (bt(Ht) − σ2 t ∇Ht ln P0|t(A(X0) = y | Ht)) dt + σtdWt,
P 0|t(A(X0) = y | Ht) =
Z 1A(x0)=y(x0)p0|t(x0|Ht)dx0
Cytaty
"Many protein design applications, such as binder or enzyme design, require scaffolding a structural motif with high precision."
"In this work, we reinterpret the conditioning problem leveraging Doob’s h-transform."
"Our main contributions are as follows: i) We derive a formal framework for conditioning diffusion processes using Doob’s h-transform."