Reproduce “Q-Diffusion: Quantizing 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
Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, Kurt Keutzer
Q-Diffusion, a post-training quantization method, effectively compresses diffusion models to 4-bit weights and 8-bit activations (W4A8) without task-specific training. It maintains image generation quality comparable to full-precision models, significantly reducing model size and improving inference efficiency across unconditional and text-guided tasks like Stable Diffusion.
Key insights
- Quantization errors in diffusion models accumulate over the iterative denoising process, demanding careful calibration to preserve output quality.
- The dynamic nature of diffusion model inputs across denoising time steps leads to significantly varying activation distributions, which static calibration methods cannot effectively capture.
- UNet shortcut connections often produce bimodal activation and weight distributions due to concatenating features with vastly different ranges, requiring specialized handling for low-bit quantization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.