HolosView R&D Dashboard

0% Setup ✓

10% Phase 0 ✓

22% Phase 1 ✓

35% Phase 2 ✓

48% Phase 3 ✓

60% Phase 4 ✓

72% Phase 5 ✓

84% Phase 6 ✦

93% Phase 7

100% Phase 8

Best Compression

730 KiB

Morton + Codebooks + Pruning · 150K→113K splats · 32.9x from PLY

Phase

6 — Product Pruning

Importance=opacity×volume beats random. Top75%: -23% size, 14% quality loss.

Pipeline

PLY → Morton → Prune → VQ → WebP

Conservative: 960 KiB. Balanced: 730 KiB. Aggressive: 530 KiB.

📊 Packing Benchmark Results

Strategy	Planes	Encoded	KiB	B/gauss	Ratio	RT
Baseline SOG-XT	9	495,112	483.5	15.5	2.57x	partial
Merged Opacity+Scales	8	494,724	483.1	15.5	2.30x	partial
Geometry Packed	7	495,734	484.1	15.5	2.02x	partial
Minimal 6-plane	7	495,734	484.1	15.5	2.02x	partial

📏 Per-Plane Entropy (Baseline)

Plane	Raw Bytes	Encoded	Ratio	Entropy	RT
active_mask	135,424	156	868.1x	0.23 b/sym	✓
means_bytes_0 (lo)	135,424	86,942	1.56x	5.88 b/sym	✓
means_bytes_1 (hi)	135,424	45,138	3.00x	5.24 b/sym	✓
opacities	135,424	24,612	5.50x	4.55 b/sym	✓
scales	135,424	78,542	1.72x	4.94 b/sym	✓
quaternions	135,424	68,460	1.98x	4.82 b/sym	⚠
f_dc (SH base)	135,424	70,994	1.91x	5.00 b/sym	✓
sh_centroids	188,160	76,510	2.46x	4.10 b/sym	✓
sh_labels	135,424	43,758	3.09x	2.86 b/sym	✓

🧬 Hypotheses Status

TESTINGH1: Adaptive RGBA channel allocation beats per-attribute planes

PENDINGH2: Block-local spatial ordering beats globally sorted layouts

PENDINGH3: Base+residual prediction improves size/decoding

PENDINGH4: Shared texture/codebook tiles reduce total bits

PENDINGH5: Shader-side quantised decoding improves startup/runtime

PENDINGH6: Product-view-aware pruning removes splats while preserving quality

PENDINGH7: Progressive model reaches interactivity sooner

PENDINGH8: Product-specific encoder/player wins on total performance

🖥️ Environment

CPU

AMD Ryzen AI Max+ 395

16 cores / 32 threads · 122 GB RAM

GPU

Radeon 8060S (RADV)

Mesa 26.0.8 · Vulkan 1.4.335 · No CUDA

Stack

Python 3.11 · Node 22 · Wrangler

NumPy · Pillow · SciPy · Cloudflare Pages/Workers/R2/D1

📂 Key Findings (Phases 0-4)

🟢 Codebooks are the big win: VQ with 64 centroids for quats (-71.5%), scales (-75.4%), SH confirmed. Total -43% vs raw WebP.
🟢 Morton ordering helps product data: -12.8% on 150K splat PLY (means_hi -76%). No gain on outdoor scenes.
🟢 Best pipeline: Morton → VQ Codebooks → WebP: 1,055 KiB for 150K splats = 22.9x compression from PLY.
🔴 Channel packing (H1) is flat: Merging channels doesn't improve WebP compression.
🔴 Predictive encoding (H3) hurts: Delta/predictive adds 8.5% overhead. WebP already has internal predictors.
🔴 WebP alpha premultiplication bug: alpha=0 corrupts RGB data. Fix: alpha=255 for all data planes.
🟡 Garden XXS is misleading: Outdoor scenes don't benefit from spatial ordering. Product data does.

📈 Compression Progression

Step	Size	Compression	Improvement
Original PLY	23.6 MB	1x	—
Quantize + Raw WebP	2,124 KiB	11.4x	baseline
+ Morton ordering	1,851 KiB	12.8x	-12.8%
+ Codebooks (SH64+Q64+S64)	1,055 KiB	22.9x	-43.2%
+ Pruning (top 75% importance)	730 KiB	32.9x	-30.8%

⏭️ Next Steps (Phase 5-6)

1. Build WebGL2 viewer to render the codebook-encoded Gaussian Splat
2. Decode centroids + labels back to full attribute arrays in shader
3. Test product-specific pruning (Phase 6) — potentially the biggest remaining win
4. Test on additional product PLYs (teddy, DeLorean)
5. Progressive loading prototype for mobile

🔬 HolosView R&D Dashboard