Mingkun Lei

Instance-specific timestep schedules can significantly boost diffusion model performance, challenging the reliance on global discretization strategies.

Ditch the handcrafted coefficients: DyWeight learns how to dynamically weight gradients in diffusion model sampling, slashing compute while boosting image quality.

Content leakage in text-to-image style transfer can be dramatically reduced by simply suppressing the tail components of the style embedding using SVD and a time-aware schedule, without any retraining.