
Motivation & Challenge: Early-constraint methods constrain intermediate features, needing one trained model per condition—costly and hard to generalize. Late-constraint freezes the diffusion network and aligns its output with conditions via a lightweight adapter, covering layout, color, structure, and shape in one solution. Results & Venues: Outperformed early-constraint baselines with better generalization to unseen conditions; timestep resampling and early stopping boosted quality. Accepted by IEEE TIP. Link: GitHub: https://github.com/AlonzoLeeeooo/LCDG https://alonzoleeeooo.github.io/LCDG/
1 like · 1 view · Sep 15, 2026
Motivation & Challenge: Early-constraint methods constrain intermediate features, needing one trained model per condition—costly and hard to generalize.