Reproduce “Post-training Quantization on Diffusion Models” 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
Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, Yan Yan
A post-training quantization method, PTQ4DM, enables 8-bit compression of Diffusion Models, achieving approximately 2x inference speedup on GPUs while maintaining or slightly improving generative quality across various datasets.
Key insights
- Activation distributions within Diffusion Models change dramatically across different denoising time-steps, requiring a dynamic calibration strategy for effective quantization.
- Calibration samples generated during the denoising process, particularly those from later steps (closer to the final image), are more effective for Post-Training Quantization (PTQ) than initial training samples.
- Post-Training Quantization can sometimes act as a regularizer, with 8-bit quantized Diffusion Models occasionally outperforming their full-precision counterparts in generative metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.