JOEY ⊹ ࣪ ˖

PHINN

DIRECTION, CG & MOTION

Welcome to my digital garden. I plant seeds haphazardly and joyfully, and see what sprouts in the spirit of exploration.

(ﾉ◕ヮ◕)ﾉ*:･ﾟ✧✿

Projects.
Fieldwork.
Info.
Play.

Insta.
Are.na.

GHOST IMPRINTS

Collective Perception
part of Sensing Tomorrow

What does a machine see when it watches a flock of pigeons lift off into the sky?

Ask it, and it will tell you. Not in confidence scores and pixel coordinates, but in the language of relationships: who's leading, who's hanging back, who's about to break away. The marks are technical: bounding boxes, network lines, heading arrows, cluster boundaries. The labels say "the straggler", "beyond reach", "still turning". It looks like data, but reads like field notes.

In the 1930s, Konrad Lorenz found that newly hatched goslings will attach themselves to the first moving thing they see, and follow it as if it were their mother. The machine imprints in a similar way. From a single frame it decides who leads and who follows, where each animal is heading, and reads everything after through that first story, until the footage forces it to change its mind. The leader becomes a follower, and the loner gets pulled back into the group. There's a strange poetry in that: a system borrowing the authority of scientific imaging to tell a story that's part evidence, part invention. Not so far from how we watch animals ourselves.

Digging into the research behind the project, we were struck by how closely this sits alongside real conservation work. Conservation X Labs uses SAM2 to follow individual animals through camera-trap footage. In Florida, where around 200 mountain lions remain, they're using it to catch the early signs of a neurological disease in the way the lions move. Their labels read "normal walking", "abnormal walking", "limping". The machine is already describing behaviour, and there, a label can be the difference between intervention and loss.

Ghost Imprints takes that same way of seeing somewhere stranger - with a light, playful touch. It doesn't claim the rigour of the science, but shares its curiosity: watching bodies move together, and asking what the movement might mean. It captures a portrait of how a collective holds together, where it fractures and who gets left behind, read by a machine still learning what any of that might mean to itself - and us, reading along with it.

____________

Sensing Tomorrow, a research brief I wrote and developed with Will Denning at Electric Theatre Collective, started from a question: as machines learn to see, move and sense the world, what happens to the way we do?

Machine perception tends towards the abstract. We reached instead for tactility, texture and craft, to ground machine sensing in human feeling and desire, and to sit in the tension where new technologies reflect back old questions about what it means to be human.

It started with a single (digitally altered) photograph of pigeons. Using Weavy and Nano Banana Pro, we asked an image model to read the relationships between the birds and draw them back over the frame as static UI overlays: bounding boxes, network lines, heading arrows. The results were beautiful, and almost entirely invented. Run the same image twice and you'd get two different flocks, with new leaders, new clusters, and lines connecting nothing. Rather than dress that up as accurate data, we leaned into it, and asked the machine for an honest, subjective reading instead.

Who is connected to whom? Which individuals are within influence range of each other?
Where are the dense social knots and where are the lonely gaps?

An image only holds one moment, though. The next question was how that perception might follow wildlife through time. It began as a dialogue with Claude Code, working out what was even possible, and quickly became a full Python workflow. We tried tracker after tracker to follow 42 pigeons across 169 frames: optical flow, a deep-learning tracker, template matching with motion prediction. Each one lost birds in its own way. What finally held was Meta's SAM3, which segments every animal in every frame, paired with a Hungarian algorithm that matches each detection to its most likely self in the next. Building this way felt intimate, like learning to see through the machine's eyes: describing what we wanted it to notice, and watching what it noticed instead.

Is the group moving as one, or fragmenting? Where are the fault lines between sub-groups?
Which individuals defy the collective direction?

Once the approach held, we rebuilt it as custom nodes in ComfyUI, and it became an evolving apparatus, retuned for every kind of camera. Pigeons filmed from an almost static viewpoint could be read by their position in the frame. Pronghorns filmed from a drone, racing across open desert, needed a system that understood the direction of the run, so that "out front" meant ahead of the herd, not the top-left of the screen.

For each clip, Claude reads five evenly spaced frames alongside the tracking data and writes a behavioural analysis of the collective, naming its characters and deciding the colours, labels and groupings. That reading splits into five lenses (movement, hierarchy, hesitation, deviation, consensus) and exports to Houdini as geometry, so every animal carries its reading with it. The footage becomes a subjective, poetic dataset that can be re-lit, re-rendered, and re-interpreted.

THE FLIGHT.

THE FEED.

THE RUN.

The machine's readings are often the best part. Watching the pronghorns, it decided they were sheep, then wrote about them with complete conviction: the herd had split, it concluded, "in velocity, not intent". Of the one falling behind, it wrote: To still be running while losing is the harder story.

(Claude has since informed me that pronghorns are decidedly not sheep. I'd believed it for months. Turns out I'd imprinted too.)

With Will Denning, Ollie Boyle, James Britton
– ETC Research Team @ Electric Theatre Collective

ROLE – Visual & conceptual research, r&d, art direction, animation, lighting & texturing, styleframes & production.
TOOLS – Figma Weave, Nano Banana Pro, Claude Code, ComfyUI, Houdini, Redshift, After Effects