Pi3 — GGUF quantizations (Lux3D)

Standalone GGUF builds of the Pi3 point-cloud reconstruction model, quantized from the canonical oxide-lab/Pi3 F32 package by Lux3D.

Each file is a complete package: resolved_config.json is embedded in the GGUF metadata, so --model-path accepts the file directly. No sibling manifest.json, checksums.json or safetensors file is required.

Files

file	dtype	size	status
`pi3-q8_0.gguf`	Q8_0, F16 fallback	0.951 GiB	recommended
`pi3-q6_k.gguf`	Q6_K, F16 fallback	0.736 GiB	usable on CUDA only
`pi3-q5_k.gguf`	Q5_K, F16 fallback	0.618 GiB	not recommended
`pi3-q4_k.gguf`	Q4_K, F16 fallback	0.507 GiB	not recommended

Canonical F32 package size: 3.57 GiB. SHA-256 and exact byte sizes are in checksums.json. Quantized matrices: 318 per variant.

Evaluation

Reference: the same CUDA backend running the canonical F32 package. Input: the three 512x512 synthetic frames in the Lux3D test assets. Metrics are drift against F32, not ground-truth accuracy: Chamfer-L1 relative to the reference bbox diagonal and symmetric F-score at 0.5%/1% of that diagonal.

variant	CD rel	F@0.5%	F@1%	F@2%
q8_0	0.166%	0.990	0.99995	1.000
q6_k	0.139%	0.985	0.99907	1.000
q5_k	0.851%	0.469	0.706	0.900
q4_k	0.813%	0.475	0.696	0.933

The same packages were re-run on every backend against that backend's own F32 reference. F@1%:

variant	CUDA	Vulkan	WGPU	CPU
q8_0	0.99995	1.000	0.9999	0.9999
q6_k	0.99907	0.966	0.919	0.895
q5_k	0.706	0.995	0.986	0.990
q4_k	0.696	0.535	0.663	0.644

Q8_0 is the only variant that is safe on all four backends. Q6_K is excellent on CUDA but loses 3–10 points of F@1% elsewhere, so use it only after validating on the target backend. Q5_K/Q4_K are backend-sensitive; keep them out of production.

Usage

# download one variant
hf download oxide-lab/Pi3-GGUF pi3-q8_0.gguf --local-dir .

# point cloud from a frame directory
lux3d-cli run pi3 --device cuda \
  --model-path pi3-q8_0.gguf \
  --source ./frames \
  --output cloud.ply

The result is an ASCII PLY with positions and RGB colors. .xyz is also accepted as a plain-text alternative. --model-path may point to the .gguf file itself or to a directory containing it.

Conversion

lux3d-cli weights quantize pi3 \
  --source /path/to/canonical-pi3/model.safetensors \
  --output pi3-q8_0.gguf \
  --dtype q8_0 --fallback-dtype f16

Only rank-2 *.weight matrices are quantized; norms, biases, embeddings, convolutions and rows that do not divide into the dtype block stay in F16. Quantized linears run through QMatMul.