TripoSG — quantized GGUF weights

GGUF builds of the Lux3D canonical TripoSG checkpoint for image-to-3D inference. Each file is standalone: the resolved config is embedded in GGUF metadata, so no model.safetensors, manifest.json, or checksums.json is needed at runtime.

Model

• Upstream: VAST-AI-Research/TripoSG (official size: 1.5B parameters in the rectified-flow transformer).
• This export serializes 1.99B parameters in total: 1.44B transformer, 304M DINOv2-large image encoder, and 243M SDF VAE.
• Source: Lux3D canonical model.safetensors, 7.40 GiB, 1349 tensors, SHA-256 68c303df01728e2b7a48c185736b6d7e1fafd7ae98618552b72e51a00da55838.
• general.architecture in every GGUF is triposg (from the config family field; the config has no top-level architecture).
• Output: triangle mesh (.obj).

Files

Uniform quantization: every rank-2 *.weight matrix uses the named dtype; norms, biases, embeddings, convolutions, and non-divisible rows stay F16. These are plain Q4_K/Q5_K/Q6_K builds, not llama.cpp _M mixed recipes.

F16 is a disk-only saving: at load time candle converts it to dense F32, so VRAM use stays close to the canonical model. Integer dtypes (Q8_0 and all Q4/Q5/Q6 variants) stay packed and run through QMatMul on CUDA/Vulkan/WGPU.

Quality

Measured on an RTX 3060 12 GiB, CUDA, seed defaults: --steps 2 --num-tokens 1024 --flash-octree-depth 5. Reference is the F32 canonical checkpoint; all variants use the same RMBG F16 matting input. Metrics are symmetric Chamfer-L1 relative to the reference bbox diagonal, F-score at 1% of that diagonal, surface voxel IoU, and normal consistency.

The input is test-assets/mesh-input-rgb.png (RGB, so the RMBG path is exercised). This is a deliberately harsh stress case for a 2-step sampler.

Two-step rectified-flow sampling is chaotic: small weight changes can move the extracted surface. On the alpha input test-assets/mesh-input.png, Q5_0/Q5_K/ Q4_K measured 0.049%/0.155%/0.253% Chamfer and Q8_0 0.026%. The Q5_K and Q4_0 outliers above are input-specific, but they are real and repeatable on that input, so do not treat a single clean run as a guarantee.

Recommendations:

• Q8_0 — default choice, near-lossless at 1.97 GiB.
• Q6_K — good compromise at 1.52 GiB.
• Q4_K — acceptable at 1.05 GiB; verify on your inputs.
• Q5_0 / Q4_1 — surprisingly strong legacy options in this test; use only if your backend handles legacy quants well.
• Q5_K / Q4_0 / Q3_K / Q2_K — not recommended for production without input-specific validation.
• F16 — use when you want a smaller file but do not want integer kernels.

Metrics are single-input numbers. For a new deployment, run the Lux3D compare_geometry.py/eval_quantization.py scripts on your own validation set.

Usage

Point --model-path either at the .gguf file or at a directory containing model.gguf:

lux3d-cli run triposg --device cuda `
  --model-path models/TripoSG-1.5B-Q8_0.gguf `
  --rmbg-model-path models/RMBG-1.4-44M-F16.gguf `
  --source input.png --output out.obj

RMBG-1.4 is a separate model and a separate GGUF because its license differs from TripoSG's MIT license. Images with a valid alpha channel skip RMBG; RGB inputs require --rmbg-model-path.

Metadata

• general.architecture = triposg
• lux3d.family = triposg
• lux3d.quantization, lux3d.fallback_dtype, lux3d.tensor_count
• lux3d.resolved_config — embedded JSON used for architecture widths.

License

The TripoSG code and weights are MIT licensed. See TripoSG-LICENSE.txt for the full text; copyright (c) 2025 VAST-AI-Research and contributors.

These GGUF files are quantized derivatives of those weights and remain under the MIT license. Keep TripoSG-LICENSE.txt with the files when redistributing. The model is provided without warranty; review the upstream terms before commercial use.

The separate RMBG-1.4 matting model is not MIT and is not covered by this license.