TRELLIS VR Equipment Optimize

Use this skill for OpenClinXR equipment generation via TRELLIS Metal → meshopt/post-opt → grade, when the goal is a learner-visible prop under station budgets, not photoreal product photography.

Hard gates (do not claim)

• Not Quest worn-headset readiness, clinical accuracy, exam equivalence, or production environmentId adoption.
• Triangle counts alone never prove visual quality — grade lit + structure (pnpm asset:model-vetting:glb-grade --glb …).
• DeepSeek models are text-only; multi-view pack generation uses Grok Imagine (image_gen / image_edit).
• Do not hyperoptimize for the metric. Meta Quest 3 class capacity is ~1.3–1.8M tris/scene (native Unity guidance). ≤40k is a multi-prop share, not the device limit. Prefer the first rung that grades as the intended prop under the preferred band; never destroy silhouette to clear 25k.

Quest 3 / WebXR budget context (research 2026-08-11)

Sources: Meta Horizon “Testing and performance analysis” (triangle + draw-call tables); Meta WebXR performance best practices + optimization workflow.

Layer	Guidance	Implication for this factory
Quest 3 / 3S scene tris (native)	1.3M–1.8M	One prop at 80–120k is tiny; full multi-actor stations need a station budget, not prop-as-scene
Quest 3 draw calls	~200–300 busy / 400–600 medium / 700–1000 light	Draw calls + materials often kill WebXR before raw tris
WebXR vs native	Small browser overhead; slow apps usually unoptimized	Optimize materials, multiview, KTX2, batching before chasing another −15k tris
Typical WebXR bottleneck	Fragment / fill-rate more than vertex	PBR everywhere, overdraw, shadows > one denser hard-surface cart
three.js / R3F / Babylon	Multiview, instancing, meshopt, maturing WebGPU↔WebXR	Headroom on CPU submit + stereo + delivery, not a free pass for raw ~1M TRELLIS

One-liner policy: Optimize until the prop is under ~60–80k and grades as the intended object; chase ≤40k only when the station’s remaining budget (actors + room + other props) requires it. Never ship raw TRELLIS. Never claim Quest readiness from triangle counts alone.

Budgets (equipment props — reframed)

Band	Tris	Use
Prop share (optional pressure)	≤ 40_000	When many props + skinned actors compete; not a quality floor
Prop preferred (stop here)	≤ 80_000	Default good enough for a single static equipment prop after grade
Prop acceptable	≤ 120_000	OK if few props / simple materials and grade prefers density
Station skeleton hard	≤ 180_000	Early partial station / few props only — not full multi-actor exam envelope
Station WebXR planning (documented)	~500k–800k	Planning envelope until worn-device profile; actors + room + equipment
Device class ceiling (Meta native Q3)	1.3M–1.8M	Whole scene; never a per-prop target

Champion selection (anti-hyperopt): among survival-ok rungs, prefer the highest triangle count still ≤ prop preferred (80k); else max ≤ acceptable (120k); else max ≤ skeleton hard (180k). Do not force 25k or “lowest under 40k” when a 60–80k rung already grades.

Optimize order (more headroom than another −15k tris): (1) draw calls / merge materials (2) cheaper secondary materials (3) KTX2 textures (4) multiview / FFR / framebuffer scale (5) then triangle count on worst offenders.

Measured findings (2026-08-10/11, ECG cart)

Input style	Multi-view bake raw tris	Post-opt floor (chain ratios)	≤180k skeleton	≤80k preferred	Visual
Photoreal product pack	~998k	~186k	miss / edge	miss	Detailed cracked/dirty clinical unit
Hard-surface / low-poly pack	~974k	~59k (chain) / 34k (high-error)	pass	pass	Boxy cart; cleaner structure; grade front may show rear

Additional:

1. Factory CLI multi-view (trellis-bake-cli.ts): all existing pack PNGs → repeated --input-image; N>1 uses sequence-concat embeddings in run_bake_isolated.py (#255).
2. Process isolation (#237): one OS process per subject or MPS OOM cascades.
3. Chain meshopt ratios plateau — after first ~0.1 cut, further 0.05…0.005 barely move tris (error bound stops simplification).
4. Photoreal inputs defeat post-opt — high-frequency detail forces dense reconstruction; optimize the prompt, not only the ladder.
5. MADR 0050 is still correct: never reject the generator on raw tris; judge after optimization. Prefer hard-surface packs so post-opt can succeed.
6. Hyperoptimize trap — forcing 25k when 60k already reads as the prop wastes silhouette for a number Quest 3 does not require per-prop.

Preferred pipeline (proven order)

1) VR hard-surface multi-view Grok packs (not photoreal)
2) pnpm factory:trellis:bake --subject <id>   # multi-view when 4 files exist
3) pnpm factory:trellis:optimize --input raw.glb --out <dir>   # high-error targets → champion
4) pnpm factory:trellis:pack --input champion.glb --out champion-meshopt.glb --compress
5) glb-grade lit + structure; three_quarter often better than front for carts
6) Record raw + best tris + paths in evidence/

meshoptimizer (zeux): already the post-opt engine (MeshoptSimplifier + optional gltfpack delivery).
factory:trellis:pack wraps gltfpack for GPU-friendly quantize/cache + -cc compression. Do not use -sa -se 1 (can zero the mesh).

Pack layout

.openclinxr/evidence/<packs-root>/<subject>/
  front.png
  side.png
  three_quarter_left.png
  three_quarter_right.png

Env: OPENCLINXR_TRELLIS_PACKS, OPENCLINXR_TRELLIS_OUT.

Hard-surface pack prompt (front) — use this first

Low-poly game-ready medical ECG monitor cart prop for WebXR / Quest.
Hard-surface stylized, NOT photoreal. Clean boxy forms only.
Wheeled base: simple rectangular platform, four chunky caster cylinders (no spokes).
Upright column: simple rectangular prism.
Main unit: rectangular box + large flat matte black screen (no glass, no waveform, no text).
Control strip: 6–8 large square button pads as raised blocks.
Ports: at most 6 large circular jacks in one row (red/blue/yellow/black), no multi-pin clutter.
NO free cables. NO logos, labels, text, brand marks.
Materials: single matte grey plastic, flat black screen, no dirt/weathering.
Studio light grey background, centered, even soft light, no floor shadow.
Maximize large flat planes for 3D reconstruction. Camera: three-quarter front elevated.

Other views: same object, only camera (side / ±40° ¾). Prefer image_edit from front for consistency.

Proven optimization technique (3 iterations)

Run after a multi-view bake. Script:
pnpm exec tsx tools/openclinxr/asset-pipeline/trellis/iterate-optimize.ts --input <raw.glb> --out <dir>

Iter	Technique	Why
1	Direct target ratio from raw with high error (error: 1) toward ~180k / ~120k / ~80k / ~60k / ~40k	Chain ratios + default error plateau; absolute targets from raw are the first fix. 25k is optional stretch only, not champion default
2	Weld (position merge) then same high-error targets	MADR 0050 step 3; removes split verts that inflate counts
3	Best quality-preserving survivor → quantize + meshopt compress delivery (retarget ≤40k only if over preferred and station share pressure)	Delivery size; do not claim topology win without grade

Record each iter’s tris, bytes, AABB volume ratio (collapse guard), and grade paths.

Constants live in iterate-optimize.ts (PROP_SHARE / PROP_PREFERRED / PROP_ACCEPTABLE / HARD) — keep skill and CLI aligned.

Grade checklist

pnpm asset:model-vetting:glb-grade --glb <raw.glb> --glb <best.glb>
# Prefer three_quarter_lit for monitor carts (front may show rear shell)

Slot	Grade
Reads as intended prop	yes/no
Flat panels / hard edges	yes/no
Interior soup / cracked screen	yes/no
Buttons/ports readable	yes/no
Tris under preferred (80k) / acceptable (120k) / skeleton hard (180k)	numbers
Hyperopt check	if ≤40k but grade worse than ≤80k sibling, prefer denser sibling

Anchors in repo

• tools/openclinxr/asset-pipeline/trellis/trellis-bake-cli.ts — multi-view factory bake
• tools/openclinxr/evidence/blender/run_bake_isolated.py — isolated Metal + multi-view cond
• tools/openclinxr/asset-pipeline/trellis/vr-postopt-ladder.ts — chain ladder (baseline)
• tools/openclinxr/asset-pipeline/trellis/iterate-optimize.ts — proven 3-iter technique (factory:trellis:optimize)
• tools/openclinxr/asset-pipeline/trellis/trellis-pack-cli.ts — gltfpack delivery (factory:trellis:pack)
• tools/openclinxr/asset-pipeline/trellis/MULTIVIEW-GROK-PACKS.md — pack operator spec
• tools/openclinxr/evidence/trellis-monitor-decimation.ts — exterior strip / deeper instruments (#250)
• Evidence examples: .openclinxr/evidence/trellis-bake/ (photoreal), trellis-bake-vr-hard/ (hard-surface)

Anti-patterns

• Photoreal “product photo” packs when budget matters
• Same-process multi-subject TRELLIS (MPS OOM)
• Endless chain ratios below 0.02 expecting VR soft targets
• Claiming VR readiness from post-opt alone without hard-surface inputs + grade
• Feeding collaged multi-view as a single image
• Treating ≤40k as Quest 3’s limit and destroying prop readability for the number
• Treating 180k station skeleton hard as a full multi-actor exam-room budget
• Picking lowest tri count under 40k when a 60–80k rung grades better

Done when a prop is “good enough to iterate in UI-XR”

• Multi-view bake exported with viewCount ≥ 2 preferred
• Proven iter report with ≥1 rung under prop preferred (80k) or documented acceptable (120k), survival ok
• Lit + structure grade reviewed (three_quarter if front is rear); hyperopt check passed
• claimScope / notEvidenceFor recorded; no readiness booleans flipped

Measured iteration results (ECG cart, VR hard-surface raw ~974k, 2026-08-11)

Iter	Technique	Target	Result tris	Survival
0	raw multi-view TRELLIS	—	973,639	ok
1	direct high-error ratio from raw	180k	179,999	ok
1	direct high-error	60k	60,000	ok
1	direct high-error	40k	39,999	ok
1	direct high-error	25k	34,443 (stretch floor)	ok
2	weld + high-error	40k	39,999	ok
2	weld + high-error	25k	34,494	ok
3	quantize from best	—	34,443	ok

Historical champion (pre anti-hyperopt selection): direct_high_error_soft25k → 34,443 tris. Still valid under prop share and preferred bands.

Policy after budget reframe: a ~60k rung (chain or high-error station target) is also a legitimate champion when grade prefers it — do not auto-prefer 34k solely because it is smaller. Site evidence still shows 34k as a measured ladder outcome, not as “Quest requires ≤40k.”

Vs chain ladder on same raw: floor ~59k with default error. High-error direct targets unlock share band when needed.

pnpm exec tsx tools/openclinxr/asset-pipeline/trellis/iterate-optimize.ts \
  --input .openclinxr/evidence/trellis-bake-vr-hard/ecg-cart/ecg-cart.glb \
  --out .openclinxr/evidence/trellis-vr-optimize-iterations/<subject>

Evidence: .openclinxr/evidence/trellis-vr-optimize-iterations/ecg-cart-vr-hard/iteration-report.json + champion.glb.