SF-V: Single Forward Video Generation Model

TL;DR

This paper introduces SF-V, a single forward pass video generation model that uses adversarial training to convert diffusion-based models into real-time video synthesizers with high quality and reduced computational costs.

Contribution

The paper presents a novel adversarial training approach to fine-tune pre-trained diffusion models for single-step, high-quality video generation, enabling real-time applications.

Findings

Achieves around 23x speedup over SVD and 6x over existing methods.

Maintains competitive video quality with significantly reduced computation.

Enables real-time video synthesis and editing.

Abstract

Diffusion-based video generation models have demonstrated remarkable success in obtaining high-fidelity videos through the iterative denoising process. However, these models require multiple denoising steps during sampling, resulting in high computational costs. In this work, we propose a novel approach to obtain single-step video generation models by leveraging adversarial training to fine-tune pre-trained video diffusion models. We show that, through the adversarial training, the multi-steps video diffusion model, i.e., Stable Video Diffusion (SVD), can be trained to perform single forward pass to synthesize high-quality videos, capturing both temporal and spatial dependencies in the video data. Extensive experiments demonstrate that our method achieves competitive generation quality of synthesized videos with significantly reduced computational overhead for the denoising process (i.e., around $23\times$ speedup compared with SVD and $6\times$ speedup compared with existing works, with even better generation quality), paving the way for real-time video synthesis and editing. More visualization results are made publicly available at https://snap-research.github.io/SF-V.