Reproduce “DGQ: Distribution-Aware Group Quantization for Text-to-Image 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
Hyogon Ryu, NaHyeon Park, Hyunjung Shim
Researchers from KAIST developed Distribution-aware Group Quantization (DGQ), a post-training method that enables efficient low-bit quantization for text-to-image diffusion models without additional fine-tuning. DGQ successfully quantizes activations to 6 bits, where prior methods largely fail, while maintaining high image quality and preserving text-image alignment, often achieving FID scores superior to full-precision models at 8-bit quantization.
Key insights
- Activation outliers are disproportionately critical for preserving image quality in T2I models and are concentrated along specific channels or pixels, rather than being uniformly distributed.
- Cross-attention scores exhibit unique distributional patterns, including a distinct peak corresponding to the `<start>` token and prompt-dependent value ranges for other tokens, which are crucial for text-image alignment.
- Tailoring quantization schemes to these specific distributional characteristics (outliers and attention patterns) is essential for effective low-bit compression in complex generative models like T2I diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.