Generative Pen-Trained Transformer
3 months ago
- #pen-plotter
- #generative-art
- #DIY
- The Generative Pen-trained Transformer (GPenT) project was conceived to create a color-changing pen plotter with a coreXY-esque kinematic system.
- Initial plans included an SVG blob generator controlled by GPT-3.5, but the project was shelved due to other commitments.
- The project was revived with the goal of creating wall-mounted art for the author's apartment using a polargraph, a vertically-mounted drawing machine.
- The polargraph design involves two motors in the top corners, belts with counterweights, and a gondola for drawing, all hidden within a wooden frame.
- Detailed assembly instructions include gondola construction with bearings and servo for pen movement, and machine assembly with wooden frame and electronics.
- Electronics setup includes stepper motors, servo, endstops, and a Raspberry Pi for control, with wiring and firmware based on Marlin and Makelangelo software.
- A web interface, plotter.local, was developed for generating and controlling plotter-compatible gcode/SVG files, accessible via Raspberry Pi.
- Sonakinatography, a visual notation system by Channa Horwitz, was adapted for the plotter to visualize poems and phrases based on syllable counts.
- An attempt was made to train a text-to-gcode diffusion model (dcode), but results were limited due to latent space optimization for pixels, not gcode.
- The final GPenT implementation uses Gemini API to generate JSON commands for plotter control, creating varied and artistic outputs.
- The project successfully produced wall art, with a gallery of favorite plots shared by the author.