Live Canvas
A line becomes an image.
The image becomes the next surface.
2026 · Interactive generative drawing system · Ongoing

Drawing as a negotiation
Live Canvas is an interactive drawing project in which a human mark becomes both image and instruction.
The hand draws on a digital canvas. After a short pause, the marks are passed into a generative process together with language. The returned image does not replace the drawing; both remain visible, allowing the artist to move between gesture and machine interpretation.
Each new line can constrain, interrupt or redirect what the system proposes. The prompt introduces another kind of authorship: not a visible mark, but a field of expectation around the image.
At what point does a line stop being a drawing and become an instruction?
The interaction loop
Gesture → Pause → Interpretation → Return → Gesture again
- 01
Gesture
A line is made directly on the iPad with the Apple Pencil. Pressure changes the weight of the mark.
- 02
Pause
When drawing stops, the newest state of the canvas and the area around the recent marks are captured.
- 03
Interpretation
The drawing becomes both an image and a spatial condition for a generative model.
- 04
Return
The generated image appears beside the drawing and can become the ground for the next gesture.
The system as material
The work uses an iPad, Apple Pencil and a private GPU server. Pencil pressure, the recent stroke area, written language, the previous generated image and controlled randomness all enter the image at different levels.
A masked image-to-image process limits change to the region around the newest marks. Sketch-based spatial conditioning can hold the drawing more firmly, while denoising controls how far the returned image may depart from the existing surface.
These settings are not hidden as neutral technical parameters. They determine how much authority remains with the hand, how much moves into language, and how much is given to the model.
- Interface
- Web canvas, Pointer Events, PWA
- Image process
- Masked img2img, ControlNet, LoRA
- Input
- Pencil gesture, pressure, text
- State
- Working prototype, ongoing
Keeping the hand present
Generative systems often ask the user to wait for a finished image. Live Canvas keeps drawing active while the image is being interpreted. New strokes can continue to arrive, but outdated requests are discarded so that the system responds only to the most recent state.
The project is not intended to automate drawing. It tests whether generation can become a responsive material—something resisted, guided and revised through repeated contact.
Current study
Ongoing · 2026Working
Pressure-sensitive drawing, manual or pause-triggered generation, local redrawing, prompt and model controls, result comparison and iterative feedback.
In development
Physical iPad documentation, measured response time, pressure-curve studies and a lighter continuous preview process.