Latest AI-generated collage

Assembled from Cutouts, Not Diffusion Pixels

Instead of generating pixels out of thin air, the single-board computer arranges physical cutouts until a neural vision model (CLIP) recognizes the subject. Click an example to inspect the physical collage metaphors:

EXHIBITION NOTES & PROVENANCE

The Collage Process: Assembled from Cutouts, Not Diffusion Pixels

Continuous algorithmic collage synthesis operating unattended on dedicated hardware since November 2021.

01 / THE AUTONOMOUS STUDIO

Assembled from Collage Cutouts, Not Diffusion Pixels

Today, most people associate “AI art” with generative diffusion models (like Midjourney or DALL-E) that hallucinate smooth, synthetic pixel grids in seconds. The 3,190+ works in Vis Plastica come from an earlier, quieter, and fundamentally different moment in computational art history.

Since November 2021, a tiny headless NVIDIA Jetson Nano has hummed continuously 24/7 in the corner of the studio running autonomous collage optimization. Instead of generating pixels out of thin air, the system starts with a blank canvas and a finite palette of distinct biological and physical cutouts — animal morphology, insect wings, feathers, shells, and botanical textures.

Drawing on Cubist collage and Dadaist chance (such as Hans Arp’s compositions), it uses Arnheim 3 (Google DeepMind’s CLIP-CLOP algorithm) to iteratively place, scale, rotate, recolor, and layer hundreds of individual patches — adjusting 2D transformations over thousands of optimization steps until a neural vision model (OpenAI CLIP) recognizes the target subject.

The Vis Plastica headless NVIDIA Jetson Nano running continuously in the studio since November 2021
Fig 1. The dedicated headless NVIDIA Jetson Nano in the studio running 24/7 unattended since November 2021.
02 / ARTISTIC METHODOLOGY

The Serendipity of Organic Cut-and-Paste

Because Vis Plastica is strictly constrained to a finite collection of real-world image cutouts, it cannot rely on photographic trickery or pixel blending. Like a human artist working with scissors, glue, and vintage nature plates, the algorithm must find unexpected visual metaphors — using a discarded perforated Nespresso pod as the iconic London 'Gherkin' in an English meadow city skyline, a headphone jack and Covid test strip as the elegant legs of Degas ballet dancers, or a stem of leaves as an ornate necklace or Elizabethan ruff on an early portrait photograph.

Operating day and night without human intervention, the Jetson Nano produces just two or three pieces a day. Because every run begins from randomized patch placements, each completed collage is a genuinely one-of-a-kind mathematical artifact — published to this archive as soon as the final render finishes.

Algorithm / Framework Arnheim 3 / CLIP-CLOP (Google DeepMind)
Academic Reference Mirowski, Banarse et al., CLIP-CLOP: CLIP-guided Collage and Photomontage (2022)
Vision Guidance Model OpenAI CLIP (Radford et al., 2021)
Hardware Architecture Dedicated Headless Computer (NVIDIA Jetson Nano, ~10W)
Operational Lifespan November 2021 – Present (5+ Years Continuous)
VIS PLASTICA

ABOUT VIS PLASTICA

Autonomous collage algorithms running 24/7 on an NVIDIA Jetson Nano using Arnheim 3 from Google DeepMind. Assembled from organic cutouts, not diffusion pixels.

3,193 Artworks
24/7 Runtime
5+ Yrs Active
STUDIO HARDWARE

The Autonomous Single-Board Computer

Vis Plastica headless NVIDIA Jetson Nano running in studio
The dedicated NVIDIA Jetson Nano humming quietly in the corner, generating 24/7 since November 2021.

Assembled from Collage Cutouts, Not Diffusion Pixels

Today, most people associate “AI art” with generative diffusion models (like Midjourney or DALL-E) that hallucinate smooth, synthetic pixel grids in seconds. The artworks in Vis Plastica come from an earlier, quieter, and fundamentally different moment in computational art history.

Since November 2021, a tiny headless NVIDIA Jetson Nano has hummed continuously 24/7 in the corner of the studio running autonomous collage optimization. Instead of generating pixels out of thin air, the system starts with a blank canvas and a finite palette of distinct biological and physical cutouts — animal morphology, insect wings, feathers, shells, and botanical textures.

Drawing on Cubist collage and Dadaist chance (such as Hans Arp’s compositions), it uses Arnheim 3 (Google DeepMind’s CLIP-CLOP algorithm) to iteratively place, scale, rotate, recolor, and layer hundreds of individual patches — adjusting 2D transformations over thousands of optimization steps until a neural vision model (OpenAI CLIP) recognizes the target subject.

Because Vis Plastica is strictly constrained to real-world image cutouts, the algorithm must find unexpected visual metaphors — using a discarded perforated Nespresso pod as the iconic London 'Gherkin' in an English meadow city skyline, a headphone jack and Covid test strip as the elegant legs of Degas ballet dancers, or a stem of leaves as an ornate necklace or Elizabethan ruff on an early portrait photograph.

"They can be a bit abstract. Sometimes good. Sometimes not so good. All of them precious."
RUNTIME DURATION 5+ Years (24/7)
TOTAL DROPS 3,193 Artworks
HARDWARE RIG NVIDIA Jetson Nano
ALGORITHM Arnheim 3 (DeepMind)