Autoregressive video diffusion models generate streaming video of arbitrary length by conditioning each new chunk on previously generated frames. However, they are prone to exposure bias, a problem that arises when a model trained on clean data is then fed its own imperfect outputs during generation.
The paper introduces "Clean Forcing" to address this issue. The method is described as drift-resistant and leverages a frozen base model. The abstract suggests that this approach helps maintain stability over long generations.
The provided abstract is partial, so specific implementation details are not available in the source. The paper is available on arXiv under the identifier 2610.25634.