A cave,a campfire,a co-designer.
A handmade miniature shelter, deconstructed into a bill of materials, re-imagined with generative AI, and re-stitched back into a tactile multimedia diorama. This is the field notebook of that process.
marker · day 0
press ambience on for the full sit-by-the-fire mood ✦
A small cave, a campfire, and a bit of help from a model.
A short note about why I made this, written for postgraduate admissions readers.
This is a small project. The artifact is a hand-built miniature cave shelter, sat on a foam base. Along the way I used some generative image tools as a sketch partner: mostly to explore variants of furniture, materials and lighting before committing them to glue and foam.
The notebook lays out, in order, the steps I actually took: a marker sketch on A4, a few rounds of AI moodboards, a bill-of-materials, then the messy days at the bench. I wanted to show how an AI tool fits into a hobby maker's workflow without pretending it did the building for me.
The next chapters walk through ideation, deconstruction, fabrication and a short reflection on what AI helped with and what it could not really touch.
A marker sketch becomes a population of variants.
Hand-drawing first, then six AI variants, then a single decision.
The marker sketch is the seed. It already contains the cave silhouette, the bench, the campfire and the back-pool with sea light. I uploaded it to an AI image tool and used it as a visual reference when asking for variants.
One scene, three benches, six material studies.
From a chosen variant to a small bill-of-materials I worked from at the bench.
reference sheet · three AI mood options · picked one
Switch the candidate
"rustic and warm. matches the campfire. picked this one."
Bill of Materials
| № | Category | Part | Material · Process |
|---|---|---|---|
| 01 | Terrain | Cave shell | EPS foam carve · paper-pulp coat · matte paint |
| 02 | Structure | Vine awning & door curtain | Real twigs · burlap · coir mat · hot-glue |
| 03 | Furniture | Tiny wooden table & bench | Popsicle sticks · bamboo skewers · tea-stain wash |
| 04 | Props | Campfire pile | Hand-stacked twigs · a pinch of ash powder |
| 05 | Flora | Moss patches & shrubs | Static grass flock · sponge clusters · matte sealer |
| 06 | Backdrop | Back-pool blue light | A printed photograph placed behind the cave hole |
Exploded View · Miniature Cave Shelter
ten parts \\\ from foam base to pebble ground \\\ assembled over four weekends
| № | Component | Difficulty | Time | |
|---|---|---|---|---|
| 01 | Foam Base Platform | Easy | 30 min | |
| 02 | Cave Structure | Hard | 4 h | |
| 03 | Door Curtain | Medium | 1 h | |
| 04 | Wooden Fence | Easy | 20 min | |
| 05 | Sleeping Mat | Medium | 1.5 h | |
| 06 | Wooden Bench | Easy | 30 min | |
| 07 | Campfire | Easy | 45 min | |
| 08 | Wood Pile | Easy | 15 min | |
| 09 | Moss Patches | Easy | 30 min | |
| 10 | Pebble Ground | Easy | 20 min | |
| Σ Total estimated time | 10 parts | ≈ 9 h 40 min | ||
A working build log, not a finished spec. Times are honest estimates from the four weekends I actually spent at the bench. Cave Structure is the only part rated Hard, mostly because of the paper-pulp drying cycles.
the BOM is what I taped to the wall before I cut the foam.
The plaster doesn't care what the prompt said.
What survives the translation from generated pixels to physical material.
The cave shell is sculpted from EPS foam, layered with paper-pulp, then washed with thinned matte emulsion. Moss comes from real static-grass flock; pebbles from a Shenzhen aquarium shop; the campfire is a hand-stacked bundle of trimmed twigs. None of this can be generated.
Every material decision was first previewed on screen, but only approved by hand — by sanding, by squinting, by photographing under three different desk lamps.
Time log
18 h sketch & prompt sweep
06 h BOM authoring
42 h sculpting + painting
04 h dressing & lighting
How would it feel under different light?
The diorama itself isn't wired. Instead, this is a small browser-side study: I overlay three CSS gradients on a photograph of the model so a viewer can imagine how the same scene might read at dawn, by the campfire, or inside a stormy night. A first sketch toward a future hardware version I would like to try.
browser-side mood study · not yet built in hardware
Trigger a scene
click each card · these are mood overlays in the browser, not real lights yet ✦
The campfire smells of glue, not of pixels.
Where the co-designer ends, and the maker begins.
The most useful thing AI gave me was options: a few quick versions of a bench, a moodboard for damp moss, a colour study before I touched paint. What it could not give me was the small physical decisions: a 3 mm gap between two twigs, a slight tilt of the table, the smell of glue at 1 am.
I think of the workflow as sketching with two hands: one hand on the marker and the foam, the other on a prompt box. The diorama is what came out of that, on a small bench, over a few weekends.
AI helped with
- · Variant exploration
- · Material moodboards
- · Colour & lighting studies
- · Some written drafts
I did by hand
- · Marker sketch & layout
- · Sourcing & sculpting
- · BOM and schedule
- · All the actual building
VII · Closing the notebook
Reflection
what the four weekends actually taught me, and what is still on the to-do list
What I Learned
- AI is good at giving me options, not at deciding scale. Picking the one that actually fits a 1:24 cave still has to happen by hand.
- Material doesn't behave like a render. Paper-pulp dries in cycles, twigs split when you bend them: the BOM had to be revised mid-build.
- Small dimensional errors compound. A 2 mm offset on the cave mouth made the back-pool photo print look obviously fake from one angle.
- Working with a model felt less like 'using a tool' and more like sketching with two hands. One on the marker, one on the prompt box.
Future Directions
wishlist- Wire the campfire and the back-pool with three small LEDs, driven by a microcontroller, so the on-screen mood study becomes a real desk-top stage.
- Try a short AR overlay (mobile camera on the diorama) that ghosts the AI concept image on top of the physical model in real time.
- Use computer vision to auto-generate a parts list from a photograph of a finished model, as an open extension of this BOM workflow.
- Carry the same loop (sketch → AI variants → BOM → bench) into a different scene, perhaps a small market stall or a temple corner.
none of these are built yet. they live in a margin of the notebook, where I keep notes for the next weekend.