Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base

Key points

• We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is 95% systematic across noise seeds.1
• Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts.1
• In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data.1
• Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.1

Sources (1)

• [1]Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 01:25 AM

  Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated.

  We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is 95% systematic across noise seeds.

Extractive summary: sentences quoted from the sources.