Firehose

This paper proposes a new training method for autoregressive video diffusion models called Self Gradient Forcing, which helps them better remember and use past information to generate future frames. Practitioners might care about this because it could lead to more realistic and stable video extrapolation.