Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models
2025/09/23 by Julien Delavande, Delavande, Julien, Régis Pierrard +4 · 13 voices · 1 citation
Business, Management and Accounting · #Private Equity and Venture Capital #cs.LG
paper · pdf · doi:10.48550/arxiv.2509.19222
openalex publication_date 2025/09/23 · openalex created_date 2025/10/16 · openalex updated_date 2026/07/28
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.
Citations
Cited by
Discussions
- arxiv.org/abs/2509.19222 [bsky, 12 points, 0 comments]
- Agreed. They are even abandoning their productivity claims now and a new paper just dropped on how video generation is exponentially more costly than image and text generation. [bsky, 5 points, 1 comments]
- Une récente étude intéressante sur le sujet, pour mesurer l’empreinte écologique de la génération de vidéo avec l’IA. C’est encore difficile à chiffrer, notamment en raison du manque de transparence d [bsky, 5 points, 2 comments]
- 1. TIL Slashdot still exists, Jesus mother-loving Christ to that. Must be some kind of money-laundering scheme. 2. I went looking for the paper and dug up the arXiv link so you don't have to. arxiv.or [bsky, 4 points, 1 comments]
- 🧵Scary paper highlighted by @ketanjoshi.co from @huggingface.co.web.brid.gy researchers. Their conclusion seemingly understates the problem. Says "a short video" = ~90Wh. But that's the second worst [bsky, 4 points, 1 comments]
- Using current text-to-video engines*, energy usage & time-to-generate scale quadratically (i.e. a 6-second clip takes 4x as long to generate, and 4x the energy draw, as a 3-second clip). So I'm curiou [bsky, 2 points, 0 comments]
- AI-generated video uses FOUR TIMES as much electricity per frame as generating a single image. arxiv.org/pdf/2509.19222 [bsky, 2 points, 0 comments]
- Characterizing the Latency and Power Regimes of Open Text-to-Video Models [hn, 1 points, 0 comments]
- I linked to the article, which linked to the original scientific paper, which explains: arxiv.org/pdf/2509.19222 [bsky, 1 points, 0 comments]
- Video Killed the #Energy Budget: Characterizing the Latency & Power Regimes of Open Text-to-Video Models "Results provide... a benchmark reference & practical insights for designing & deploying more s [bsky, 0 points, 0 comments]
- AI video generation is exponentially more energy intensive when time or resolution doubles. This. Is. Insane. arxiv.org/pdf/2509.19222 [bsky, 0 points, 0 comments]
- And the icing on the cake is new research showing the longer the AI video, the worse the proportional impact on the environment: arxiv.org/pdf/2509.19222 🤦♂️ [bsky, 0 points, 0 comments]
- "Video Killed the Energy Budget" 😂 (tl;dr: AI video creation runs real hot) #mlsky arxiv.org/abs/2509.19222 [bsky, 0 points, 0 comments]
Related