𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡丨Late-Constraint Controllable Visual Generation

AI Researcher & First Author

𝐌𝐨𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧 & 𝐂𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞: Condition-specific diffusion controls struggle to generalize across guidance types. For this long-standing issue, we propose LaCon, a generalizable method that aligns external conditions with diffusion features and applies late constraints to steer generation.

𝐑𝐞𝐬𝐮𝐥𝐭𝐬 & 𝐕𝐞𝐧𝐮𝐞𝐬: Flexible control across modalities, granularities and diffusion backbones. Published in IEEE Transactions on Image Processing, 2026.

𝐋𝐢𝐧𝐤：https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11360571