Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout
Autoregressive (AR) video diffusion models have shown great potential in real-time video generation. Recent methods distill pretrained bidirectional video diffusion models into causal AR students through Distribution Matching Distillation (DMD), but the generated videos often suffer from over-saturation and over-smoothing issues, resulting in limited visual quality and realism.
Reading in Plain English mode — toggle below to switch
Autoregressive (AR) video diffusion models have shown great potential in real-time video generation. Recent methods distill pretrained bidirectional video diffusion models into causal AR students through Distribution Matching Distillation (DMD), but the generated videos often suffer from over-saturation and over-smoothing issues, resulting in limited visual quality and realism.
This research addresses key challenges in autoregressive video generation, specifically improving visual quality and realism by mitigating mode collapse. For ML engineers, understanding these distillation and mode-collapse mitigation techniques is crucial for developing efficient and high-quality real-time video generation systems.