VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation
Researchers have introduced Velocity Decomposition and Estimation (VDE), a novel training-free method to accelerate rectified flow models used in generative tasks. VDE decomposes the model's velocity into components that are estimated based on temporal predictability and directional stability, moving away from traditional caching techniques. This approach aims to improve inference speed with minimal impact on visual quality, as demonstrated by experiments on image and video generation. AI
IMPACT Accelerates inference for generative AI models, potentially enabling wider adoption in real-time applications.