Train Short, Inference Long: Training-free Horizon Extension for Autoregressive Video Generation

TL;DR

FLEX is a training-free inference framework that enhances autoregressive video models to generate longer videos by adaptively interpolating frequency components and injecting dynamic priors, significantly improving long-term video synthesis.

Contribution

We introduce FLEX, a novel training-free method combining frequency-aware modulation and dynamic priors to extend video generation horizons without retraining models.

Findings

Outperforms state-of-the-art models at 6x extrapolation (30s)

Matches long-video fine-tuned baselines at 12x scale (60s)

Supports 4-minute dynamic video synthesis

Abstract

Autoregressive video diffusion models have emerged as a scalable paradigm for long video generation. However, they often suffer from severe extrapolation failure, where rapid error accumulation leads to significant temporal degradation when extending beyond training horizons. We identify that this failure primarily stems from the spectral bias of 3D positional embeddings and the lack of dynamic priors in noise sampling. To address these issues, we propose FLEX (Frequency-aware Length EXtension), a training-free inference-time framework that bridges the gap between short-term training and long-term inference. FLEX introduces Frequency-aware RoPE Modulation to adaptively interpolate under-trained low-frequency components while extrapolating high-frequency ones to preserve multi-scale temporal discriminability. This is integrated with Antiphase Noise Sampling (ANS) to inject high-frequency dynamic priors and Inference-only Attention Sink to anchor global structure. Extensive evaluations on VBench demonstrate that FLEX significantly outperforms state-of-the-art models at 6x extrapolation (30s duration) and matches the performance of long-video fine-tuned baselines at 12x scale (60s duration). As a plug-and-play augmentation, FLEX seamlessly integrates into existing inference pipelines for horizon extension. It effectively pushes the generation limits of models such as LongLive, supporting consistent and dynamic video synthesis at a 4-minute scale. Project page is available at https://ga-lee.github.io/FLEX_demo.