Reproduce “Accelerating Diffusion via Hybrid Data-Pipeline Parallelism Based on Conditional Guidance Scheduling” 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

Euisoo Jung, Byunghyun Kim, Hyunjin Kim, Seonghye Cho, Jae-Gil Lee

A hybrid data-pipeline parallelism framework accelerates conditional diffusion model inference, achieving a 2.31x latency reduction for Stable Diffusion XL and 2.07x for Stable Diffusion 3.0 on two GPUs. This approach maintains or improves generation quality by leveraging condition-based partitioning and adaptive parallelism switching based on denoising discrepancy.

Key insights

  • The inherent two-path structure of Classifier-Free Guidance provides a natural, artifact-free way to partition data across GPUs, preserving global image consistency.
  • The denoising discrepancy between conditional and unconditional noise estimates follows a U-shaped curve, allowing for adaptive scheduling of pipeline parallelism to optimize speed and quality.
  • By strategically switching between data-only and hybrid data-pipeline parallelism stages, the framework minimizes communication overhead and error propagation, leading to superlinear speed-ups and improved image fidelity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.