Retro Photo Craze Revives Questions About the Energy Behind AI-Generated Images
The latest social-media trip to the 1980s takes only a photograph and a prompt.
Users are feeding contemporary pictures into generative AI tools and asking for versions that resemble vintage studio portraits: period hairstyles, old-fashioned clothing, warm lighting, film grain and the visual language of family albums or classic cinema.
The results have spread rapidly across Instagram and other platforms, helped by celebrities, politicians and public figures joining the trend.
What appears on a phone as a quick piece of nostalgia, however, is produced by computing infrastructure that consumes real-world resources. As the number of AI-generated portraits grows, the trend is bringing renewed attention to the electricity and cooling required to produce generative images at scale.
The important question is not whether one retro portrait carries an enormous environmental cost. Available research does not support such a simple conclusion.
Scale is where the calculation changes.
A Generated Portrait Is Not Just Another Photo Filter
The distinction begins with what happens after a user submits an image.
A conventional photo filter generally modifies an existing picture using predetermined adjustments. Generative AI can perform a substantially more complex task, interpreting both an image and instructions before constructing a new visual output.
That processing typically relies on specialised computing hardware, including GPUs operating inside data centres.
For an individual user, the process can feel almost frictionless. Upload a selfie, describe the desired decade or aesthetic and wait for the result.
But users rarely stop at one attempt. They may change a hairstyle, outfit, background or prompt, generating several alternatives before choosing the version they want to post.
A viral trend therefore produces more than a large number of finished pictures. It can produce many more generation requests behind them.
There Is No Reliable Universal Figure for One AI Image
Putting a precise environmental price on those requests is difficult.
Research comparing 17 image-generation models found that electricity consumption could vary by as much as 46 times between the systems tested. The reported range ran from about 0.086 watt-hours to 4.08 watt-hours per image, depending on the model and testing conditions.
That spread is important.
It means there is no scientifically useful single number that can be applied to every AI-generated portrait. Model architecture, hardware, image resolution, quantisation and other technical choices can materially alter the amount of computation required.
Nor is there a verified total for the electricity consumed specifically by the current retro-photo trend.
Without knowing how many images have been generated, which models produced them and where the computing took place, any attempt to assign the craze a precise global energy or carbon footprint would rely heavily on assumptions.
What researchers can measure is the broader relationship: image generation requires computation, and additional computation requires electricity.
Viral Popularity Turns Small Requests Into a Scaling Problem
The environmental significance of generative AI becomes clearer when individual requests are considered collectively.
Retro portraits have spread through several variations, including looks inspired by vintage Bollywood, old family photography and regional aesthetics. Indian politicians including Commerce and Industry Minister Piyush Goyal and Parliamentary Affairs Minister Kiren Rijiju have participated in the trend, alongside actors and other public figures.
That visibility encourages further participation and, in turn, more image generation.
The pattern is familiar from earlier AI crazes. A visual style catches attention, users experiment with it, and millions of computing requests can follow within a short period.
This does not mean the trend itself is a major driver of global electricity consumption. Data centres support far larger workloads, including cloud computing, streaming, business applications, search and an expanding range of AI services.
The retro-photo craze is better understood as a small, highly visible example of a much bigger change in computing behaviour.
Generative AI is becoming something consumers use casually.
AI Arrives as Data-Centre Electricity Demand Is Already Rising
That shift matters because AI adoption is accelerating alongside broader growth in data-centre electricity consumption.
AI models now support far more than image creation. They generate and analyse text, write software, produce audio and video, assist with search and increasingly appear inside workplace applications and consumer services.
Each workload has a different computational profile.
A still image may require relatively modest resources compared with a complex AI-generated video. But frequent, low-cost interactions can still become significant when multiplied across a very large user base.
The wide efficiency gap between image-generation models also points to another side of the issue. Rising AI usage does not necessarily mean energy consumption has to increase at exactly the same rate.
If developers can produce comparable results with more efficient models and hardware, the energy required for each interaction can fall even as usage expands.
The Resource Question Extends Beyond Electricity
Electricity is only part of the infrastructure behind an AI-generated image.
The chips carrying out the computation produce heat. Data centres need cooling systems to keep that hardware within safe operating temperatures, and some of those systems consume water.
How much depends heavily on the facility.
Location, climate, cooling technology and the source of electricity can all affect the water footprint associated with a computing workload. Electricity production itself may also require water, adding another variable to attempts to calculate an AI service’s total resource use.
For that reason, claims that every AI image consumes a fixed amount of water deserve the same caution as universal electricity estimates.
A figure derived from one facility or one model cannot automatically be applied to every image created through every consumer AI service.
AI Video Shows How Quickly Computing Requirements Can Rise
The contrast becomes more pronounced as generative systems move beyond still pictures.
Research published in 2026 by the United Nations University Institute for Water, Environment and Health estimated that generating a typical AI image could require energy equivalent to running a 10-watt LED bulb for roughly 17 minutes.
For a high-complexity AI-generated video, its comparison rose to around 42 hours of operation for the same bulb.
Such comparisons are useful for communicating differences in computational intensity, but they should not be treated as fixed measurements for every AI platform or prompt.
Different models, hardware configurations and generation settings produce different results.
The larger point is that the next phase of consumer generative AI is likely to involve increasingly rich media. If AI video becomes as routine as AI image generation, efficiency gains will become more consequential.
The Bigger Issue Is Efficiency, Not One Viral Craze
The popularity of retro portraits can make the environmental debate seem like a choice between entertainment and sustainability. That framing is too narrow.
Generative AI is already being applied across science, medicine, education, software development, business and creative work. The same underlying infrastructure can support both consequential applications and disposable social-media experiments.
The more useful question is how efficiently that infrastructure operates as demand grows.
New generations of chips can reduce the energy required for particular workloads. Developers can optimise models. Data centres can improve cooling efficiency, while greater use of low-carbon electricity can reduce emissions associated with the power they consume.
Greater transparency would also help.
Consumers generally have little way of knowing whether one AI service uses substantially more energy than another to produce a similar result. Even researchers can struggle to obtain detailed information about proprietary models and the data centres running them.
That makes viral trends such as the current ’80s-photo craze useful less as evidence of a specific environmental crisis than as a glimpse of what mass-market AI looks like.
A single portrait may be a tiny computing task.
Millions of people generating, rejecting and regenerating images are something different: a reminder that the future environmental footprint of AI will be determined not just by what the technology can do, but by how efficiently it can do it at enormous scale.
