Skip to main content

420 AI physical card background image generations in an Agroecology course for young people?

How does that fit together? “Watt”, if all educational comic illustrations were created on just one consumer graphic card on a self hosted open-source AI model? Running smoothly on 17 Watt energy consumption in idle work mode and a maximum of 435 Watt, when 10 concurrent students generated at the same time from multiple devices in a classroom (63 degree Celsius heat development during peak performance.



This refurbished system can run offline on a lower mid end consumer Gaming PC/Laptop graphic card, which could run solar/wind powered, solo or in a local grid for the 1h image generation. 

Thats the Beauty of AI evolution today. We are down-scaling at the grassroot up!
Last Monday. Mission briefing:

  • Ethical and responsible use of a low energy consuming, Open-Source AI Image Generation tool for food system game card creation in school pilot agroecology workshops in East Africa.
  • Easy to use interface + has to work on (older) tablets, PCs, smartphones + high availability in rural regions + easy cloud sync backup of daily results and sharing with multi-stakeholder team on multiple continents
  • has to be cost efficient
  • Live Monitoring Telemetry for serving the students from remote backoffice field support, ensuring the digital field tools keep running smoothly as much as those in the real gardens.

I developed the Web App from Open-Source building blocks (quick recipe below) in 16 non-stop hours (the first 8 hours figuring out a general model and direction, slightly panicking at the end, because I had no happy solution, yet), then rented an on-demand server pod for 0,60$ per minute to make the App available for school pilot 1 a couple hours later. Between school days we only payed 0.02 $ per minute for the sleep mode of our App server and GPU.

All running great last week – until it became 15 concurrent users this week, each generating 3 image variations with their manually written prompt and the coolest filter out there.. then our GPU “burned” into the offline at 65 degree Celsius and our Ubuntu Linux got hickups.. 😀 The stress test, web developers are in a strange way happy about (while we heavily sweat).

Quick recipe list:

  • Rented short term server with
    1 NVIDIA RTX4090 on runpod.io
  • Open-Source model Flux2 by bfl.ai (Black Forest Labs, Germany) (Hey fellow Cyber Druids and neighbors of ancient forests…HUGE FAN! Thank you!!)
  • invoke.ai for a highly customizable, easy to use front end with multiple model choices. Big Wow. Even with 15 concurrent device users in a single user setup. Whutt? Yes, turned out to be a feature 😀
  • Self hosted Matomo for privacy focused Web-App analytics and live monitoring https://matomo.org/

See Results on https://www.fsysgame.org