Autoregressive video generation improved with long-term look ahead
Prospective Forcing: Learning to Look Ahead for Autoregressive Long Video Diffusion
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Generating long videos by predicting them one piece at a time can cause problems because the model only learns to predict the very next part. The authors propose a method called Prospective Forcing that teaches the model to also consider multiple future parts ahead. This helps the video stay consistent and clear over longer durations. It speeds up video creation by sometimes skipping parts of the process when confident in future predictions. The system can also generate multiple guesses and choose the best one to improve accuracy.
What this means in practice
- •For video production teams: Produce longer videos with improved quality and consistency by training models to predict multiple future frames at once.
- •For game developers: Generate extended video sequences efficiently by using predicted future chunks to reduce rendering time during content creation.
Authors
Yifan Shen, Yewei Xia, Gongxu Luo, Peiyuan Zhu, Boyang Sun, Zhenhao Chen, Guangyi Liu, Guangyi Chen, Kun Zhang
Abstract
Autoregressive video generators synthesize long videos through chunk-wise rollouts, yet they are typically trained to predict only the immediate future. This limited prediction horizon makes it challenging to maintain video quality over extended rollouts. To address this problem, we introduce Prospective Forcing, a multi-head prediction framework that injects long-range predictive signals into autoregressive video representations. Built on a rolling-window generator, Prospective Forcing attaches lightweight, offset-specific heads that share the same conditioning information and independently predict chunks at different future temporal offsets. By extending supervision beyond the next chunk, these heads encourage the shared history representation to retain information useful for longer-horizon prediction, supporting temporally consistent generation. At inference time, we accelerate sampling by using accepted future-chunk predictions to skip full-window backbone executions. We equip the head with learnable embeddings to generate multiple candidates for the next chunk, and a learned verifier either selects a candidate or rejects all and falls back to standard autoregressive generation.