InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO

2025-05-23

AI Summary

To address semantic inconsistency across scenes, weak text-prompt alignment, and prohibitive computational overhead from context growth in autoregressive long-video generation, this paper proposes a training-free, inference-time dynamic context selection framework. Our method introduces: (1) a learnable context selection policy optimized via Group Relative Policy Optimization (GRPO); (2) a multi-objective hybrid reward function jointly optimizing semantic alignment, cross-scene consistency, and artifact suppression; and (3) a dynamic top-K semantic filtering mechanism enabling efficient inference under fixed computational budgets. Evaluated on the CsVBench benchmark and the EPS event prompting dataset, our approach extends video length by up to 9× while significantly improving cross-scene consistency and semantic fidelity. The implementation is publicly available.

Technology Category

• Computer Vision: Computational Photography, Image & Video Synthesis
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Recent advances in text-to-video generation, particularly with autoregressive models, have enabled the synthesis of high-quality videos depicting individual scenes. However, extending these models to generate long, cross-scene videos remains a significant challenge. As the context length grows during autoregressive decoding, computational costs rise sharply, and the model's ability to maintain consistency and adhere to evolving textual prompts deteriorates. We introduce InfLVG, an inference-time framework that enables coherent long video generation without requiring additional long-form video data. InfLVG leverages a learnable context selection policy, optimized via Group Relative Policy Optimization (GRPO), to dynamically identify and retain the most semantically relevant context throughout the generation process. Instead of accumulating the entire generation history, the policy ranks and selects the top-$K$ most contextually relevant tokens, allowing the model to maintain a fixed computational budget while preserving content consistency and prompt alignment. To optimize the policy, we design a hybrid reward function that jointly captures semantic alignment, cross-scene consistency, and artifact reduction. To benchmark performance, we introduce the Cross-scene Video Benchmark (CsVBench) along with an Event Prompt Set (EPS) that simulates complex multi-scene transitions involving shared subjects and varied actions/backgrounds. Experimental results show that InfLVG can extend video length by up to 9$ imes$, achieving strong consistency and semantic fidelity across scenes. Our code is available at https://github.com/MAPLE-AIGC/InfLVG.

Problem

• Enables coherent long video generation without extra data
• Reduces computational costs via dynamic context selection
• Maintains consistency and prompt alignment across scenes

Innovation

• Uses GRPO for dynamic context selection
• Maintains fixed computational budget via top-K tokens
• Hybrid reward optimizes alignment and consistency