Reproduce “DiTAS: Quantizing Diffusion Transformers via Enhanced Activation Smoothing” 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
Zhenyuan Dong, Sai Qian Zhang
Researchers at NYU developed DiTAS, an efficient data-free Post-Training Quantization (PTQ) framework for Diffusion Transformers (DiTs), achieving near full-precision image generation quality at challenging low bitwidths. DiTAS attained an FID-10K of 6.86 for W4A8 on ImageNet 256x256, surpassing previous state-of-the-art methods and making DiTs deployable on resource-constrained hardware.
Key insights
- Effective low-bit quantization for DiTs requires addressing both the dynamic temporal variability and persistent channel-wise outliers in activation distributions.
- Aggregating outlier statistics across all denoising time steps, rather than a single snapshot, is critical for robust smoothing in diffusion models.
- Combining refined activation smoothing with an error-compensating LoRA module for weights allows for maintaining high generative fidelity even at 4-bit weight precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.