Building Little Nis: From Hike Photo to WebGPU Paper World
Translating a photograph of Thomas Dambo's reclaimed-wood forest giant beside Holly Pond in Bernheim Forest into a tactile cut-paper and Kubelka-Munk watercolor diorama.
The whole project started from a single photograph I took on a trail in Clermont, Kentucky: Thomas Dambo's reclaimed-wood giant Little Nis crouching at the edge of Holly Pond in Bernheim Forest, staring at his upside-down double in leaf-stained autumn water. I wanted to convert that photograph into a live 3D experience that felt crafted out of folded rag paper and mineral watercolor washes — without photogrammetry meshes, baked textures, or screen-space cartoon filters.
The hardest lesson was learning what not to enclose. An early attempt to hide the island's outer boundary with a 6-sheet deckle plinth and dense overhead canopy choked out the open Kentucky sky and turned the pond muddy, forcing a full overnight reset. Putting the raw hike photo back beside the viewport made the fix obvious: watercolor needed to live inside each individual cut-paper shingle, bark facet, and autumn leaf via 2-flux Kubelka-Munk glazes ( over #f2ebd9 Washi), paired with an explicit scale.y = -1 reflection assembly beneath the water plane and zero debug chrome.
After stripping out every slider for a five-plate scroll-choreographed trail folio and running a 1:1 release pass against the photograph — continuous screen-space Washi fibers, pale-grey beech and sycamore crowns, salvaged plank face and 38-slat collar, and clear olive-ochre water holding 95 fps at Retina 2880×1800 — Little Nis shipped live at little-nis.andrewvoirol.com.
Timeline
I took a field photo from a hike in Kentucky and started converting it into a cut-paper and watercolor diorama in Three.js r187 WebGPU — without a single photogrammetry scan or image texture. Sparked by a Three.js WebGPURenderer sketch demo, I wanted to see if TSL (three/tsl) node materials could build the scene purely from procedural paper craft and 2-flux Kubelka-Munk pigment glazes over warm unbleached Washi (#f2ebd9). Instead of flat digital alpha blending, pigment film thickness enters directly against the paper backing , paired with multi-octave irrational-angle FBM cold-press tooth (clamp(dot(N, L), 0.94, 1.06)). By late afternoon the first prototype compiled cleanly and rendered a crouching wooden shingle figure above a capillary-rippled pond at 60+ fps, proving the optical stack worked before any anatomical sculpting began.
My first attempt to frame the clearing backfired and buried the openness that made the sketch work. The initial island floated in empty space with a hard cutoff, so I stacked the ground into a 6-sheet torn deckle plinth, wrapped a dense leaf canopy overhead, and added folded-paper tree trunks. Technically it closed the horizon, but visually the heavy overhead foliage choked out the sky gaps above the trees and muddied the pond reflection underneath. After burning two hours nudging canopy density and wrap angles that still looked cramped, I called the turn a miss, committed the checkpoint, killed the dev server, and shut down for the night to reset.
Putting the raw hike photo back on screen exposed why the second pass regressed: watercolor belongs inside each cut-paper piece, not smeared over the camera. Looking closely at the reference photo, the woodland needed open sky breathing through sparse autumn branches, and the pond needed a clear upside-down guardian hanging in the water. I thinned the forest canopy, rebuilt the pond around an explicit scale.y = -1 inverted reflection assembly perturbed by radial capillary wave dispersion, and rewrote paperCraftShader.ts so every individual cedar shingle, bark facet, and leaf cutout carries its own wet-in-wet Kubelka-Munk glaze variation (vWashSeed) and evaporative edge darkening (fuchidori). Once each folded element looked cut from hand-painted watercolor rag paper before assembly, the scene clicked back into focus.
Once the diorama held together optically, leaving shader sliders on screen felt like taping a mixing console onto a storybook. The sculpture in my photo is Little Nis — Thomas Dambo's 2019 reclaimed-wood forest giant crouching beside Holly Pond in Bernheim Arboretum and Research Forest in Clermont, Kentucky. I stripped out every pill bar and parameter slider, locked the Autumn Tannin #f2ebd9 sheet, and turned the page into a five-stop scroll-choreographed trail folio (Trailhead plus Plates I–IV). Scrolling interpolates camera pose, FOV, and camera.setViewOffset so the diorama stays framed in the right two-thirds of the sheet beside each folklore slip, while sweeping the cursor blooms pigment along the paper path and clicking Holly Pond casts ripples across his reflection.
Right before release, a 1:1 audit against my original Bernheim photo caught five places where procedural defaults had drifted away from the real place. Side by side with the photograph, five gaps jumped out: the canvas cleared to a flat hex color instead of continuous fibrous paper, the trees drooped like green umbrellas instead of tall pale-grey Kentucky beech and sycamore trunks, Little Nis lacked his flat-plank brow and radial scrap-wood collar, a 0.44 sky-sheen term washed Holly Pond out in milky haze, and a generic grey boulder sat where Bernheim's wooden trail plaque actually stands. I wired a full-viewport screen-space scene.backgroundNode Washi shader (trimmed from 12 to 4 noise samples so Retina 2880×1800 holds 95 fps / 10.6 ms), rebuilt the plank cranium, downcast eyes, and 38-slat collar (20,860 triangles), dropped pond sky-sheen to 0.16 so the inverted reflection reads clearly through olive-ochre water (R_inf = (0.15, 0.19, 0.08)), and planted the angled wooden plaque on the right bank before deploying to little-nis.andrewvoirol.com.