Art Generator
The weather, rendered daily, by a hand that is not the same hand twice.
Generative artAtmospheric dataTrained style models
Art. renders daily generative art from live atmospheric data. Weather patterns, satellite palettes, and environmental signals are converted into original artworks, on a rotation of seventeen artists, in PNG and print-quality output.
Most generative art is prompted, which means its subject is ultimately the person writing the prompt. Here the subject is the sky over a specific place on a specific morning: pressure, humidity, wind, and precipitation are the input, and nobody chooses them. The work is figurative in an unusual sense, in that it is a picture of something that was actually there.
The seventeen artists are the second half of the argument. Each one holds a fixed sensibility, its own palette, mark, and restraint, so the same weather resolves seventeen different ways and none of them is the house style. Rotation makes the point that a system with a single aesthetic is not an artist, it is a filter.
How it is built
The input is measurement, not mood: hourly Open-Meteo observations for the day's locations, and Sentinel-2 scenes sampled for colour, so a palette comes off actual ground and cloud rather than a designer's swatch. Nine of the seventeen artists render through their own weights, FLUX LoRAs trained on curated sets of a single artist's canvases, as few as fourteen where copyright limits what can be sourced.
A vision model then scores every finished piece for fidelity to that artist, which is what drives the work: the artist rendering worst is the one who gets trained next.
- Data
- Open-Meteo hourly, Sentinel-2 via Copernicus
- Renderer
- FLUX 1.1 Pro, with per-artist LoRAs
- Training
- 9 LoRAs, 14 to 66 canvases each
- Evaluation
- Vision model scores artist fidelity, 1 to 10
- Pipeline
- Step Functions and Python Lambdas on AWS, daily