Reproduce “MosaicDiff: Training-free Structural Pruning for Diffusion Model Acceleration Reflecting Pretraining Dynamics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bowei Guo, Shengkun Tang, Cong Zeng, Zhiqiang Shen
Researchers at MBZUAI developed MosaicDiff, a training-free structural pruning method that accelerates diffusion model sampling by adapting pruning strategies to the varying learning dynamics observed during model pretraining. This approach significantly reduces computational costs while maintaining or improving generation quality, outperforming prior methods on DiT-XL/2 and SDXL-base-1.0 models.
Key insights
- The pretraining dynamics of diffusion models exhibit a distinct 'slow-fast-slow' multi-stage learning phase, where different timesteps contribute unequally to the final image quality.
- Quantifying the importance of sampling steps, by considering both the rate of change in Mean Squared Error (MSE) and Signal-to-Noise Ratio (SNR), enables effective adaptive pruning.
- Training-free structural pruning, when guided by pretraining dynamics and using carefully constructed, stage-specific calibration data, can achieve superior acceleration without compromising generation quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.