Free Video-LLM: Prompt-guided Visual Perception for Efficient Training-free Video LLMs

TL;DR

This work tackles the high computational cost of training-free video LLMs by introducing Free Video-LLM, a prompt-guided visual perception framework that decouples temporal and spatial processing to selectively attend to content relevant to the input prompt. By employing prompt-guided temporal sampling and RoI-based spatial cropping, the method significantly reduces visual token counts while maintaining competitive accuracy on four video QA benchmarks. Experiments show substantial efficiency gains, including faster pre-fill latency and higher tokens-per-second, with results that rival or exceed existing training-free approaches. The approach demonstrates a practical pathway to scalable, resource-efficient video understanding suitable for real-time applications.

Abstract

Vision-language large models have achieved remarkable success in various multi-modal tasks, yet applying them to video understanding remains challenging due to the inherent complexity and computational demands of video data. While training-based video-LLMs deliver high performance, they often require substantial resources for training and inference. Conversely, training-free approaches offer a more efficient alternative by adapting pre-trained image-LLMs models for video tasks without additional training, but they face inference efficiency bottlenecks due to the large number of visual tokens generated from video frames. In this work, we present a novel prompt-guided visual perception framework (abbreviated as Free Video-LLM) for efficient inference of training-free video LLMs. The proposed framework decouples spatial-temporal dimension and performs temporal frame sampling and spatial RoI cropping respectively based on task-specific prompts. Our method effectively reduces the number of visual tokens while maintaining high performance across multiple video question-answering benchmarks. Extensive experiments demonstrate that our approach achieves competitive results with significantly fewer tokens, offering an optimal trade-off between accuracy and computational efficiency compared to state-of-the-art video LLMs. The code will be available at https://github.com/contrastive/FreeVideoLLM.

Figures (2)

• Figure 1: Illustration of the proposed Free Video-LLM with prompt-guided visual perception.
• Figure 2: Example video question answering results on MSVD-QA benchmark.