Title: FlowC2S: Bridging Current and Succeeding Frames for Rapid, Memory-Conserving Video Continuation

Abstract: We present FlowC2S, a new approach designed to produce video continuations that are both swift and memory-efficient. This technique involves fine-tuning a pre-trained text-to-video flow model to establish a vector field linking current video segments with the subsequent ones. The methodology relies on two critical innovations. First, it employs inherent optimal couplings by using temporally neighboring video chunks during the training phase as a practical substitute for genuine optimal couplings, which leads to more linear flow paths. Second, it integrates target inversion, a process that injects the inverted latent representation of the target chunk into the input to enhance correspondence accuracy and boost visual quality. Unlike conventional methods that combine current frames with noise to generate continuations, FlowC2S flows directly from current to succeeding frames. This strategy cuts the model’s input dimensionality in half. When fine-tuned from either LTXV or Wan, our method achieves superior quantitative performance in FID and FVD metrics, requiring as few as five neural function evaluations.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC