LIFT and PLACE: A Simple, Stable, and Effective Knowledge Distillation Framework for Lightweight Diffusion Models

Researchers propose LIFT and PLACE, a knowledge distillation framework that enables stable training of extremely lightweight diffusion models by decomposing the teacher's complex denoising process into coarse and fine stages with spatially adaptive guidance. The method achieves stable convergence even at extreme compression ratios (1.6% of teacher size) where conventional distillation fails, with potential applications across image generation, latent diffusion, and flow-based models.

The paper addresses a fundamental bottleneck in deploying diffusion models at scale: the gap between large, capable teacher models and the compressed student models needed for practical deployment. Traditional knowledge distillation struggles because student networks cannot effectively learn from the teacher's highly complex denoising trajectories, leading to training instability and quality degradation at high compression ratios. LIFT and PLACE solve this by introducing a coarse-to-fine learning strategy that breaks down the distillation objective into manageable sequential stages, allowing students to build foundational understanding before tackling harder refinement tasks.

This work extends beyond incremental optimization—it demonstrates architectural robustness across diverse settings including U-Net and DiT backbones, conditional and unconditional generation, and emerging models like MMDiT used in Stable Diffusion 3. The extreme compression scenario (1.3M parameters versus the teacher's ~80M) is particularly noteworthy, as conventional methods degrade catastrophically while LIFT maintains stable training with competitive FID scores of 15.73 versus 50-200+ for baseline approaches.

For the AI and machine learning community, this framework directly impacts the accessibility and deployment efficiency of generative models. Practitioners can now confidently compress diffusion models to edge-friendly sizes without sacrificing quality, accelerating adoption in resource-constrained environments like mobile devices and embedded systems. The spatially adaptive coefficient estimation component suggests the method generalizes well to heterogeneous error distributions, indicating potential applicability to other model compression tasks beyond diffusion.

• →LIFT and PLACE enables stable knowledge distillation of diffusion models at extreme compression ratios (1.6% of teacher size) where conventional methods fail
• →Coarse-to-fine decomposition allows students to learn complex teacher behaviors progressively, preventing training collapse
• →Framework generalizes across architectures (U-Net, DiT), modalities (image, latent), and emerging models including flow-based designs
• →Achieves FID score of 15.73 under extreme compression compared to 50-200+ degradation with standard distillation
• →Spatially adaptive guidance via PLACE addresses non-uniform error distributions, improving local generation quality