[Track #1]: paper

Landing page!
3D manipulation sliders and and their performance on targeted platforms!
AI co-pilot to make adjusting your favorite 3D objects to your needs easy!
Wire-frame view of object and workspace!
View of object and workspace!
Dragging and dropping desired 3D objects for editing!
Edited 3D object exported, prepped for target platform, and easily imported to blender!

paper · 3D optimization for everyone

Inspiration

Shipping a 3D model to a mobile game today means installing Blender, learning the Decimate Modifier, configuring texture compression, and iterating until something works; or paying a tech artist $80/hour. Tools like Blender, Maya, and AutoCAD are powerful but built for experts. We wanted to compress that workflow into three sliders and a button.

What it does

Drop in any GLB, pick a target; Mobile, VR, Mid PC, or High PC, and walk away with a platform-ready file in seconds. paper gives you:

Three real-time sliders: topology, texture quality, and vertex precision
Live platform compatibility rings showing headroom on each tier
An AI co-pilot (Llama 3.3 70B via Groq) that diagnoses your asset and recommends a preset for the chosen platform
A silent visual-integrity check (windowed SSIM) that warns the AI about hole-like damage when sliders go too aggressive
One-click GLB export with a SHA-256 integrity hash

Everything runs client-side; your asset never leaves your browser.

How we built it

Vite + React + TypeScript shell, three.js + react-three-fiber for the WebGL viewer, meshoptimizer WASM for decimation, and three.js's GLTFExporter for binary GLB output. Groq powers live AI recommendations with Zod schema validation; a hand-tuned fallback cache keeps the demo bulletproof if Groq rate-limits mid-pitch.

The optimization pipeline is a four-stage transform: LOD swap → texture downsample → vertex quantization → re-analyze. Each fires on every slider change. Visual fidelity is computed by rendering pristine + current to a 192² offscreen canvas and running windowed SSIM:

$$ \text{score} = 0.6\,\overline{\mathrm{SSIM}} + 0.4\,\overline{\mathrm{SSIM}}_{\text{worst }10\%} $$

Worst-decile weighting catches localized damage that global SSIM averages away.

Challenges

SSIM globally averaged is useless for catching holes. A baseball with a visible tear scored 0.96. Switching to windowed SSIM with worst-decile weighting fixed it.
Polyhaven scans hit a topology floor because photogrammetry meshes have disconnected fragments. Aggressive mergeVertices pre-pass + iterative simplifySloppy fallback unlocked further reduction.
LOD swap silently broke vertex quantization; mesh.geometry references change, dropping the quantized state. Fix: combined the two effects into one ordered pipeline.
Per-asset budgets aren't scene budgets. Initial draw-call caps were meaningless for single-material assets. We replaced them with vertex-buffer memory, which actually scales.
Live demos can't fail. A 12-entry fallback cache (3 assets × 4 platforms) with a one-click toggle keeps the AI panel populated even with zero tokens.

What we learned

The hard part of 3D optimization isn't the algorithms; meshoptimizer, Draco, and three.js's exporters all exist as libraries. The hard part is making them legible to someone who isn't already a tech artist. That's UX work, not graphics work, and it's where the win is.