Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation
The paper’s core claim is that video distillation works better when the teacher sees only the same past context the student had at generation time.
CMD replaces full-clip, bidirectional teacher scoring with a causal teacher that cannot use future frames or controls. It also scores targets against the student-generated prefix that actually led to them, then perturbs weak early prefixes to stabilize training. The authors report state-of-the-art aggregate results among autoregressive methods on short- and long-video benchmarks, plus stronger adherence to changing camera controls. HF Daily Papers' note
CMD replaces full-clip, bidirectional teacher scoring with a causal teacher that cannot use future frames or controls. It also scores targets against the student-generated prefix that actually led to them, then perturbs weak early prefixes to stabilize training. The authors report state-of-the-art aggregate results among autoregressive methods on short- and long-video benchmarks, plus stronger adherence to changing camera controls. HF Daily Papers' note
score 5