TripoSR — GGUF quantizations (Lux3D)

Standalone GGUF builds of the TripoSR image-to-mesh model, quantized from the canonical oxide-lab/TripoSR 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

Canonical F32 package size: 1.56 GiB. SHA-256 and exact byte sizes are in checksums.json. Q8_0 quantizes 244 matrices; the K-quants use 256-wide blocks, so nine matrices whose row length is not divisible by 256 stay in F16 (235 quantized matrices).

Evaluation

Reference: the same CUDA backend running the canonical F32 package. Input: the burger test image from the Lux3D test assets, marching cubes at resolution 256 and threshold 25.0. Metrics are drift against F32, not ground-truth accuracy: Chamfer-L1 relative to the reference bbox diagonal, F-score, surface voxel IoU and normal consistency.

TripoSR is the most quantization-tolerant family in this group: every variant keeps F@1% ≥ 0.9997 on CUDA, Vulkan, WGPU and CPU. Q8_0 cross-backend F@1%: 0.9999 CUDA / 1.000 Vulkan / 1.000 WGPU / 0.9999 CPU. Q4_K still loses surface topology (voxel IoU 0.66–0.68 across backends versus 0.92–0.94 at Q8_0), so prefer Q6_K/Q8_0 when topology matters.

Usage

# download
hf download oxide-lab/TripoSR-GGUF triposr-q8_0.gguf --local-dir .

# mesh from a single image (GLB by default)
lux3d-cli run triposr --device cuda \
  --model-path triposr-q8_0.gguf \
  --source ./input.png \
  --output mesh.glb

# legacy OBJ / other mesh formats
lux3d-cli run triposr --device cuda \
  --model-path triposr-q8_0.gguf \
  --source ./input.png \
  --output mesh.obj

Mesh output dispatches on the extension: .glb (primary), .obj, .ply, .stl. TripoSR also emits per-vertex colors, stored as COLOR_0 in GLB and RGB in PLY/OBJ; STL drops colors.

Conversion

lux3d-cli weights quantize triposr \
  --source /path/to/canonical-triposr/model.safetensors \
  --output triposr-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.

License

TripoSR code and weights are distributed under the MIT License; see LICENSE. Model weights remain subject to the upstream license.