Video World Models · 2026-10-01

LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation

Byoungwoo Park, Jaemoo Choi, Juho Lee, Yongxin Chen

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

LongTake trains an autoregressive video diffusion model with Long-Horizon Teacher Forcing (TF) on curated real long videos. Long-Horizon TF supervises later frames using long ground-truth prefixes, strengthening direct initialization for self-rollout distribution matching distillation (DMD) under the same 5s DMD setup. A second-stage Hybrid DMD reuses the Long-Horizon AR teacher to conditionally supervise later frames, achieving the highest dynamic degree at 30s and 60s while maintaining strong quality.

Why it matters

The authors address failures in long-horizon AR video generation where rollouts become near-static or lose visual quality due to limited short-clip supervision. By using real long videos for extended teacher forcing and then reusing that teacher during student self-rollouts, LongTake improves sustained dynamics over long durations and places on the Pareto front of dynamic degree vs aesthetic quality.

Method

Long-Horizon TF: construct conditioning KV-cache from long real-video prefixes and directly supervise prediction of later target frames beyond the short training horizon.
Two-stage training: initialize the student for self-rollout DMD directly from Long-Horizon TF weights, avoiding the intermediate few-step distillation initialization stage used in standard pipelines.
Hybrid DMD: reuse the Long-Horizon AR teacher to conditionally supervise later frames of the student self-rollout while retaining bidirectional joint supervision over the initial window.

Limitation

On 30s and 60s rollouts, the authors evaluate dynamic degree against aesthetic quality and VBench quality; however, the paper notes that stagnation can be easy to miss because consistency metrics can favor low-dynamic videos.

Abstract (from arXiv)

World models, game simulators, and long-take video creation require coherent scene evolution and sustained dynamics over extended durations. Autoregressive (AR) video diffusion provides a natural framework for long-horizon generation, yet extended rollouts often become near-static or lose visual quality. We hypothesize that these failures reflect the limited guidance provided by short-video supervision on how ongoing scene dynamics develops over longer durations. This motivates us to introduce LongTake, a two-stage training pipeline built around Long-Horizon Teacher Forcing (TF) on curated real long videos. Long-Horizon TF trains the AR model to predict later frames conditioned on long ground-truth video prefixes, extending direct supervision beyond the short training horizon. This supervision is designed to help the model sustain dynamics and preserve visual quality during long-horizon generation. Our central finding is that this training stage strengthens direct initialization for distribution matching distillation (DMD) under student self-rollout, without the intermediate few-step distillation stage used in standard pipelines. Under the same five-second DMD training setup, our initialization yields substantially higher dynamic degree than short horizon TF initialization on 30-second rollouts at comparable aesthetic quality, and surpasses the evaluated baselines in both measures. Hybrid DMD further reuses this teacher to extend supervision to later frames of the self-rollout while retaining bidirectional joint supervision over the initial window. On long-horizon self-rollouts, LongTake lies on the Pareto front of dynamic degree and aesthetic quality, and Hybrid DMD attains the highest dynamic degree among evaluated methods at both 30s and 60s.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-09-29 · Indexed by Pub-AI: 2026-10-01 UTC

preprint