# How do I cut the impact of AI image and video generation?

> Why AI images and video use far more energy than text, and practical ways to generate fewer, shorter, smaller outputs without losing what you need.

Source: https://greenai.guide/how-do-i/cut-impact-of-ai-image-and-video-generation/  
Updated: 2026-11-19

## In short

Generate fewer images and videos, plan the prompt before you start, keep video short and at the resolution you need, and reuse good results. Image generation used over a thousand times more energy per request than text classification in one study, and a short AI video can cost far more again.

## Steps

1. Decide whether a stock image, photo or simple diagram would do the job.
2. Write a detailed prompt before generating, so fewer attempts are needed.
3. Ask for one or two images, not a large batch of variations.
4. Keep videos short and no higher resolution than the place they will be shown.
5. Edit or refine a near-miss instead of regenerating from scratch.
6. Save and reuse images that work.

**Images and video are the heaviest everyday AI tasks, so they are where restraint pays off most.** In tests by
Luccioni, Jernite and Strubell, generating images used on average about 2.9 kWh per 1,000 images, against
0.002 kWh per 1,000 text classifications: more than a thousand times as much per request.

## Why is video so much heavier?

**Because every frame has to be generated, and cost grows quickly with length and resolution.** Delavande,
Pierrard and Luccioni measured open text-to-video models and found energy grew quadratically with the video's
spatial and temporal size. One short video from one model used about 90 Wh, which they put at roughly 30 times an
image and 2,000 times a text answer. The spread between models was large: from about 0.14 Wh to over 415 Wh.

These are measurements of open models on research hardware. Commercial services may differ, and most do not
publish figures.

## How do I generate less?

- **Plan first.** Describe subject, style, composition and format in one prompt.
- **Ask for fewer variations.** Four images cost about four times one.
- **Go short and small for video.** A few seconds at the size it will be viewed is usually enough.
- **Use fewer steps** where a tool lets you choose quality settings: the same study found halving the generation
  steps roughly halved the energy.
- **Reuse.** Keep a folder of images you can use again.

## What makes the biggest difference?

Avoiding the "slot machine" habit of generating dozens of options for fun. Treat media generation like printing in
colour: fine when it is needed, wasteful by default.

Read more in [use AI efficiently](/practice/use-ai-efficiently/), and teachers can try the "Text, image or video?"
activity in our [lesson ideas](/schools/lesson-ideas/).

## Sources

- [Power Hungry Processing: Watts Driving the Cost of AI Deployment?](https://arxiv.org/abs/2311.16863), Luccioni, Jernite and Strubell, ACM FAccT (2024)
- [Video killed the energy budget: characterizing the latency and power regimes of open text-to-video models](https://arxiv.org/abs/2509.19222), Delavande, Pierrard and Luccioni (arXiv 2509.19222) (2025)
