Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models

2025-09-23

AI Summary

Text-to-video (T2V) generation suffers from high latency and energy consumption, yet its resource usage patterns remain poorly characterized. This work presents the first systematic evaluation of runtime latency and power consumption across six open-source T2V models—including WAN2.1-T2V—combining analytical modeling with fine-grained empirical measurements. We uncover fundamental scaling laws: energy consumption grows quadratically with spatial resolution and temporal length, and linearly with denoising steps. Building on these insights, we introduce the first energy-efficiency benchmark suite for open T2V models and propose a predictive, computation-constrained analytical model. Our findings provide quantitative foundations and practical design guidelines for algorithmic optimization, hardware co-design, and sustainable deployment of green generative video systems.

Technology Category

• Computer Vision: Video Understanding & Activity Analysis
• Natural Language Processing: Generation
• Machine Learning: Efficient ML / Green AI

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Energy management for devices in mobile Web and WoT environments
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Recent advances in text-to-video (T2V) generation have enabled the creation of high-fidelity, temporally coherent clips from natural language prompts. Yet these systems come with significant computational costs, and their energy demands remain poorly understood. In this paper, we present a systematic study of the latency and energy consumption of state-of-the-art open-source T2V models. We first develop a compute-bound analytical model that predicts scaling laws with respect to spatial resolution, temporal length, and denoising steps. We then validate these predictions through fine-grained experiments on WAN2.1-T2V, showing quadratic growth with spatial and temporal dimensions, and linear scaling with the number of denoising steps. Finally, we extend our analysis to six diverse T2V models, comparing their runtime and energy profiles under default settings. Our results provide both a benchmark reference and practical insights for designing and deploying more sustainable generative video systems.

Problem

• Characterizing latency and energy consumption of text-to-video models
• Developing analytical model for scaling laws prediction
• Benchmarking runtime and energy profiles across diverse models

Innovation

• Developed compute-bound analytical model for scaling laws
• Validated predictions through fine-grained WAN2.1-T2V experiments
• Extended analysis to six diverse T2V models