The authors held a noisy video state fixed and compared the model’s denoising velocity, its prediction for how noise should be removed, under clean versus previously generated histories.
Small frozen-model corrector curbs drift in long autoregressive video generation
Clean Forcing learns systematic errors caused by generated history, reducing degradation over 50-second rollouts without retraining the base model.
Big Tech
Wenqing Wang · Joonghyuk Shin · Jonathan Tremblay · Chan Hee Song · Yun Fu
Seoul National University · Northeastern University · NVIDIA
Research Digest··3 min read
Wang et al.
§