Reproduce “TR-DQ: Time-Rotation Diffusion Quantization” 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
Yihua Shao, Deyang Lin, Fanhu Zeng, Minxi Yan, Muyang Zhang, Siyu Chen, Yuxuan Fan, Ziyang Yan and 5 more
TR-DQ introduces a post-training quantization framework for diffusion models, leveraging time-dependent rotation matrices and attention-sharing to reduce inference memory and latency. It achieves superior generative quality at 4-bit and 8-bit quantization for both image and video tasks, outperforming existing methods.
Key insights
- Activation distributions in diffusion models exhibit significant, dynamic variations across time-steps, necessitating adaptive, time-dependent quantization strategies for optimal performance.
- "Massive outliers" in activations are a critical barrier for effective diffusion model quantization, which can be mitigated by sophisticated rotation-based balancing strategies.
- Classifier-Free Guidance introduces substantial computational redundancy through similar attention mechanisms in conditional and unconditional paths, offering a viable target for compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.