Hydro(power)

02 — 2026 · generative · water, optical flow, diffusion

Hydro(power) is a single-channel video in which every movement is borrowed from water. Footage of rivers, waterfalls and hot springs is reduced to optical flow: a record of where, and how quickly, the water moves in each frame, with the image itself discarded. These fields of motion are then used to steer a diffusion model. The model supplies the picture; the water supplies only the current that carries it.

What the model is asked to imagine is water at each of the scales at which it shapes the world: bacterial colonies blooming across a wet plate, lichen spreading over stone, the mineral terraces of a hot spring, deltas and braided rivers seen from orbit. None of these places exists. Each is a statistical proposal, assembled from the model’s training images and set in motion by a river it has never seen. Across five sections joined by slow dissolves, the image passes from the microscopic to the planetary without a cut.

The title names a second current. The data centres that train and run models such as this one consume vast quantities of water to cool their processors, often drawing on rivers and aquifers already under strain. In Hydro(power) that dependency is made visible and turned around: water no longer cools the machine but moves it, lending its motion to images the model could not otherwise animate. The work does not stand outside the system it describes. Every frame was made at the same cost.

Year
2026
Type
generative
Model
Diffusion model, optical flow, ControlNet, frame interpolation
Latent
64×36×4 · stable diffusion 1.5 vae (kl-f8)
Credits
Single-channel video generated by custom software, silent. 7 min 47 s.