One-Forcing: Towards Stable One-Step Autoregressive Video Generation

One-Forcing introduces a stable one-step autoregressive video generation method by combining the DMD objective with an auxiliary GAN loss.

It achieves state-of-the-art one-step results on VBench, beating prior one-step methods and nearing stronger multi-step systems.

The method uses about one-third of the training cost of the chunkwise model, making low-latency video generation more practical.