LIFT and PLACE: Efficient Knowledge Distillation for Lightweight Diffusion Models
Researchers propose LIFT and PLACE, a coarse-to-fine knowledge distillation framework for lightweight diffusion models. LIFT decomposes the objective into coarse alignment and fine refinement, while PLACE addresses spatially non-uniform errors with locally adaptive guidance. Experiments show effectiveness across diffusion spaces, backbones, tasks, and datasets, extending to flow-based models.
Key facts
- LIFT stands for LInear FiTting-based distillation.
- PLACE stands for Piecewise Local Adaptive Coefficient Estimation.
- The framework addresses the challenge of teacher network's complex denoising process.
- LIFT decomposes the objective into coarse alignment and fine refinement.
- PLACE partitions outputs into error-based groups for locally adaptive guidance.
- Effective across image and latent diffusion spaces.
- Effective across U-Net and DiT backbones.
- Effective for unconditional and conditional tasks.
Entities
Institutions
- arXiv