Generating a video with AI carries an energy cost nobody has been measuring consistently - until now.
A new paper from arXiv introduces a bidirectional framework for estimating the energy consumption of text-to-video and text-to-audio-video models using only observable output parameters: resolution, duration, and architectural principles. No access to model weights, internal code, or proprietary implementation details is required. The researchers validated the approach across six open-source models ranging from 8.3 billion to 27 billion parameters on three GPU configurations, landing below 3% mean absolute percentage error across all tested architectures. The math decomposes each model's energy profile into quadratic and linear terms whose coefficients map directly to architectural complexity.
The implications are practical. Right now, comparing the carbon footprint of competing video generation systems is nearly impossible - labs don't publish energy figures, and replicating inference at scale to measure power draw is expensive. A standardized, outside-in estimation method gives researchers, regulators, and procurement teams a way to benchmark sustainability without vendor cooperation. That matters as text-to-video models scale up and the compute bills - and grid demand - grow with them.
The framework works in both directions: forward to predict consumption from known parameters, backward to infer architectural behavior from observed inference times. It won't settle debates about which lab is greenest, but it gives critics and auditors a defensible methodology - something the industry has been conspicuously uninterested in providing itself.