TokenTrim: Inference-Time Token Pruning for Autoregressive Long Video Generation

TL;DR

TokenTrim introduces an inference-time token pruning method that identifies and removes unstable latent tokens during autoregressive long video generation, significantly enhancing temporal consistency without altering the model or training process.

Contribution

The paper proposes a novel inference-time token pruning technique that mitigates temporal drift in long video generation by removing corrupted latent tokens, without changing the model architecture or training.

Findings

Improves long-horizon temporal consistency in video generation.

Effectively identifies and removes unstable latent tokens during inference.

Enhances video quality without retraining or modifying the original model.

Abstract

Auto-regressive video generation enables long video synthesis by iteratively conditioning each new batch of frames on previously generated content. However, recent work has shown that such pipelines suffer from severe temporal drift, where errors accumulate and amplify over long horizons. We hypothesize that this drift does not primarily stem from insufficient model capacity, but rather from inference-time error propagation. Specifically, we contend that drift arises from the uncontrolled reuse of corrupted latent conditioning tokens during auto-regressive inference. To correct this accumulation of errors, we propose a simple, inference-time method that mitigates temporal drift by identifying and removing unstable latent tokens before they are reused for conditioning. For this purpose, we define unstable tokens as latent tokens whose representations deviate significantly from those of the previously generated batch, indicating potential corruption or semantic drift. By explicitly removing corrupted latent tokens from the auto-regressive context, rather than modifying entire spatial regions or model parameters, our method prevents unreliable latent information from influencing future generation steps. As a result, it significantly improves long-horizon temporal consistency without modifying the model architecture, training procedure, or leaving latent space.