Co-Evolving Paths and Flows via Path-Flow Alignment
The paper proposes training the path and the flow together, then adds stochastic regularization to stop the learned path from collapsing through narrow intermediate states.
The authors say plain path-flow alignment can overfit: the loss falls, but sample quality gets worse. They trace that failure to low-entropy bottlenecks in the probability path. Their regularizer hides some source information from the path network while keeping exact endpoints, giving the training marginals an entropy floor. On ImageNet-256x256 with SiT backbones, they report consistent FID gains without changing inference architecture or sampler. ArXiv · AI/CL/LG's note
The authors say plain path-flow alignment can overfit: the loss falls, but sample quality gets worse. They trace that failure to low-entropy bottlenecks in the probability path. Their regularizer hides some source information from the path network while keeping exact endpoints, giving the training marginals an entropy floor. On ImageNet-256x256 with SiT backbones, they report consistent FID gains without changing inference architecture or sampler. ArXiv · AI/CL/LG's note
score 5