Reproduce “GPD: Guided Progressive Distillation for Fast and High-Quality Video Generation” 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
Xiao Liang, Yunzhu Zhang, Linchao Zhu
Researchers from Zhejiang University developed Guided Progressive Distillation (GPD), an acceleration framework for video diffusion models that achieves an 8x speedup by reducing inference steps from 48 to 6, while surpassing the base model and other state-of-the-art methods on the VBench benchmark with a total score of 84.04%. This approach integrates online teacher refinement, progressive training, and high-frequency preservation to maintain visual quality and temporal consistency.
Key insights
- Adaptive, online guidance from a teacher model, based on the student's intermediate predictions, effectively addresses distribution shifts and leads to more robust and accurate trajectory learning.
- A progressive training approach that gradually increases the student's prediction step size stabilizes the distillation process and enables efficient learning of complex, non-linear diffusion trajectories.
- Explicitly penalizing the loss of high-frequency components in both spatial and temporal dimensions, particularly during later diffusion stages, is critical for preserving visual fidelity and dynamic realism in accelerated video generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.