What it is
A two-pass training scheme for autoregressive video diffusion that lets future losses train how self-generated history is written into KV memory for long video rollout. Unlike Self Forcing, which trains on self-rollouts but freezes historical KV cache writes, SGF replays a sampled rollout step in a parallel second pass so future losses supervise clean-timestep KV memory writing without full rollout backprop.
Key results
- Matches Self Forcing on 5s VBench short-horizon evaluation
- Improves over Self Forcing at 60s native long-video extrapolation
- Improves over Self Forcing at 240s and extrapolates to several minutes
Long-horizon video consistency depends not just on reading past cache, but on training how self-generated history is written into causal memory.