A field guide to making an artist’s work responsive, surprising and unmistakably their own.
Creative coding, material simulation and AI offer very different ways to make an image move. This guide maps the possibilities—and where the evidence ends.
Start with the recommendation ↓
Begin with the artist’s materials. Offer one clear action. Make the before and after recognizable. Let a trace remain—or give the visitor something to take away.
This is our design recommendation, not a measured finding. It combines source fidelity with agency: an image becomes more than a spectacle when the visitor can understand what changed and why. Preserve the original alongside any interpretation, and let the artist approve the source and the treatment.
A procedural shader follows authored rules. An AI model uses trained weights. Both can be creative, but they are not interchangeable descriptions.
Each starts from a source image and gives the visitor a different kind of agency. These are digital interpretations, not new original paintings or evidence of physical pigment.
13,300 source-texture brush quads pull apart and return. Its trajectories are analytic and stateless; there is no compute-physics simulation. The surprise is seeing a recognizable image become a field of marks.
Explore the assembly ↗A 640² retained ping-pong feedback field begins with The Listener. Press and drag to deposit colour, then export a PNG. This is feedback advection, not a fluid solver or AI.
Make a pigment print ↗A 144² displaced, depth-tested mesh turns source colour gradients into an imagined surface. Its height is inferred; it is not a scan or measurement of the painting.
Explore the terrain ↗Together: recognition and release, persistent authorship, and an alternative point of view. They test different relationships to the same source rather than three versions of one effect.
These examples document what their makers describe. Their relevance to Kochu is an editorial interpretation, not proof that the same effect will work here.
For Myriad (Tulips) and Mosaic Virus, Ridler photographed and labelled 10,000 tulips. The selection and naming are part of the work. Start with an approved collection, not an assumption that five paintings make a style model. Artist’s process ↗
Learning to See uses custom neural networks to interpret a live camera feed through learned examples. A visible input and response can make the system’s point of view legible. Artist’s project ↗
Hobbs describes how spacing, curve length and distortion shape a flow field. The useful lesson is to author constraints deliberately, not simply add noise. Artist’s essay ↗
Unsupervised used an institutional image collection and trained models; the listed work is non-interactive. Its custom latent-space browser suggests a separate way to explore alternatives, not a claim about that work’s interaction. Studio project ↗
Artificial Natural History uses neural networks to imagine organisms in a fictive catalogue. Here, any categories for the paintings should be authored or approved by the artist. Artist’s project ↗
Fresco documents simulated watercolour bloom and oil mixing in its Live Brushes. Material behaviour can itself be the event; the local feedback study should not be described as Fresco’s physics engine. Official documentation ↗
Search the 23 directions or filter by readiness. “Build now” is a suggested browser direction, not a claim that every technique has already been built.
Use this as a shortlist, not a compatibility guarantee. Verify current versions, model weights, source-image permissions and the terms of every dependency before production. “Rights check” is a prompt for review, not legal advice.
| Tool / primary source | Useful for | Data & compute | Rights scope to check |
|---|
More approved photographs and artist-authored crops would support a richer index; consistent multi-view captures would be needed to investigate a real-space reconstruction. A custom neural cellular automaton needs trained weights and tested inference, while Gaussian splatting requires an appropriate capture and reconstruction pipeline. Neither is an MVP for five flat photos. A generated or inferred depth map is not a measurement of paint thickness. NCA research ↗ Original splatting paper ↗
Primary links throughout were checked on 31 October 2026 in the accompanying research. Tool documentation changes; revisit it before implementation. The recommendations are hypotheses, not visitor-test results. No AI generation or model training is claimed for these studies.
Save the complete field guide as one HTML file ↓
The three implemented studies use the pinned native vgpu 0.6.0 API. They render image-derived geometry and retained texture state; they do not run an AI model. The other tools below are a research shortlist, with individual implementation work still required.
The vault’s older ComfyUI long guide is withdrawn/historical; its old speed, hardware and pricing claims are not relied upon. Older p5 material contains broad support and performance claims; use current official documentation. The 3D-tour note is an analogy from another domain, not evidence about these paintings.
← Return to all studies