Reproduce “SegQuant: A Semantics-Aware and Generalizable Quantization Framework for 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
Jiaji Zhang, Ruichao Sun, Hailiang Zhao, Jiaju Wu, Peng Chen, Hao Li, Yuying Liu, Kingsum Chow and 2 more
SegQuant is a post-training quantization framework for diffusion models, improving deployment efficiency and image fidelity by employing a semantics-aware graph-based segmentation for linear layers and a hardware-native polarity preservation technique for activations. The framework achieves superior image quality metrics, such as an FID of 22.85 for FLUX.1-DiT, while ensuring compatibility with AI compilers.
Key insights
- Linear layers in diffusion models, especially DiT-based architectures, often concatenate semantically distinct input features, requiring segment-wise quantization for optimal performance.
- Polarity-asymmetric activation functions in modern diffusion models contain crucial low-magnitude negative values that are poorly resolved by conventional quantization, necessitating dedicated resolution for negative and non-negative ranges.
- Quantization strategies must be derived from static computation graph analysis, avoiding runtime-dynamic data or manual, architecture-specific rules, to ensure compatibility with modern AI compilers and automated deployment pipelines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.