Long-Horizon Continuation
in Autoregressive Video Diffusion
2
Coherent scene evolution and sustained dynamics
over extended durations.
01 / METHOD
LongTake learns from curated real long videos, then trains a few-step generator under student self-rollout. Long-Horizon TF provides a strong direct initialization for self-rollout DMD, without a separate few-step distillation stage.
Short clips provide no direct supervision for prediction beyond their training horizon. Long-Horizon TF addresses this teacher gap by learning to predict later frames from long ground-truth video prefixes with a bounded KV-cache.
At inference, the KV-cache comes from generated frames rather than ground-truth videos. Self-rollout DMD addresses this student gap, starting directly from the Long-Horizon TF model. Optional Hybrid DMD reuses the AR teacher for conditional supervision of later frames, while retaining bidirectional joint supervision over the initial window.
02 / TEACHER INITIALIZATION
Short TF, Short TF + CD, and Long TF under the same 5s self-rollout DMD training, evaluated with 30s rollouts.
All three use joint DMD only (λ = 0), isolating teacher initialization from the optional conditional guidance in Hybrid DMD.
03 / COMPARISONS
Compare 30s and 60s autoregressive rollouts with the same prompt and seed.
λ controls the conditional DMD weight. We use λ = 0.2 for the main comparisons and gallery.
04 / VIDEO GALLERY
Selected qualitative examples of long-horizon continuation, coherent scene evolution, and sustained dynamics.