Reproduce “VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation” on OpenResearch

About this paper

Junwen Tan, Jinglin Liang, Hongyuan Chen, Shuangping Huang

Researchers at South China University of Technology developed Velocity Decomposition and Estimation (VDE), a training-free method to accelerate rectified flow models by dynamically estimating model output based on velocity component regularities. The approach achieves up to a 3.22x speedup while maintaining high visual fidelity, including a 52.2% reduction in LPIPS compared to other training-free baselines on the Qwen-Image model.

Key insights

• The scalar coefficients of the parallel and orthogonal velocity components exhibit smooth, locally linear temporal evolution, making them predictable through extrapolation.
• The unit vector defining the orthogonal direction of the velocity remains highly stable over short time intervals, allowing for its reuse across multiple steps with minimal error.
• By explicitly estimating velocity components from the current input, VDE effectively mitigates the 'cache-input mismatch' problem inherent in prior caching-based methods.

Summary by alphaXiv.