One-Forcing: One-Step Autoregressive Video Generation
Researchers propose One-Forcing, a method for one-step autoregressive video generation that combines DMD with GAN loss to improve quality and reduce latency. Existing few-step methods degrade severely when reduced to one step, producing weak dynamics or blurry frames. One-Forcing achieves a VBench score of 83.76, setting a new state-of-the-art. The approach addresses the challenge of efficient real-time interactive video generation.
Key facts
- One-Forcing is a one-step autoregressive video generation method.
- It augments DMD objective with auxiliary GAN loss.
- Existing few-step methods degrade in one-step setting.
- One-Forcing achieves VBench total score of 83.76.
- It addresses quality degradation in one-step generation.
- Trajectory-style consistency distillation produces weak dynamics.
- DMD-based approaches like Self-Forcing yield blurry frames.
- The method targets real-time interactive video generation.
Entities
—