oxide-lab/VGGT-1B-GGUF

VGGT-1B — quantized GGUF weights

GGUF builds of the Lux3D canonical VGGT-1B checkpoint for multi-view geometry: point maps, depth, camera extrinsics/intrinsics, and confidence. Each file is standalone: the resolved config is embedded in GGUF metadata.

Model

• Upstream: facebookresearch/vggt, weights facebook/VGGT-1B.
• 1.26B parameters in this export (the name keeps Meta's VGGT-1B label).
• Components: DINOv2 ViT-L/14 backbone, 24 frame + 24 global aggregator blocks, 4-iteration camera head, DPT point/depth heads. The query-point track head is not part of this inference slice.
• Source: Lux3D canonical model.safetensors, 4.68 GiB, 1797 tensors.
• general.architecture = vggt (the config has neither architecture nor family, so the model name is used); lux3d.family = vggt.
• Output: confidence-filtered point cloud (.ply) plus depth and cameras.

Files

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

file	weights	size	quantized matrices	fallback F16
VGGT-1B-F16.gguf	F16	2.34 GiB	387	1410
VGGT-1B-Q8_0.gguf	Q8_0	1.33 GiB	384	1413
VGGT-1B-Q6_K.gguf	Q6_K	1.09 GiB	331	1466
VGGT-1B-Q5_K.gguf	Q5_K	0.95 GiB	331	1466
VGGT-1B-Q4_K.gguf	Q4_K	0.82 GiB	331	1466
VGGT-1B-Q5_0.gguf	Q5_0 (legacy)	0.93 GiB	384	1413
VGGT-1B-Q4_0.gguf	Q4_0 (legacy)	0.80 GiB	384	1413
VGGT-1B-Q4_1.gguf	Q4_1 (legacy)	0.86 GiB	384	1413
VGGT-1B-Q5_1.gguf	Q5_1 (legacy)	1.00 GiB	384	1413
VGGT-1B-Q3_K.gguf	Q3_K	0.68 GiB	331	1466
VGGT-1B-Q2_K.gguf	Q2_K	0.57 GiB	331	1466

F16 is a disk-only saving: candle converts it to dense F32 at load time, so VRAM use stays close to the canonical model. Integer dtypes stay packed and use QMatMul on CUDA/Vulkan/WGPU.

Quality

Measured on an RTX 3060 12 GiB, CUDA, 3 frames from test-assets/pi3-frames, against the F32 canonical checkpoint. Camera-pose AUC uses the two consecutive relative poses. Point-map and depth PSNR/AbsRel are computed from the same tensors exported with --dump-npz.

VGGT is pose-sensitive; geometry metrics alone are not enough.

file	pointmap PSNR	depth PSNR	Chamfer rel	F@1%	AUC@5	AUC@15	AUC@30
F16	71.18 dB	75.76 dB	0.043%	1.000	1.0	1.0	1.0
Q8_0	53.15 dB	48.08 dB	0.274%	0.999	1.0	1.0	1.0
Q6_K	45.65 dB	44.78 dB	0.571%	0.902	1.0	1.0	1.0
Q5_K	34.74 dB	35.93 dB	1.537%	0.488	0.0	1.0	1.0
Q4_K	24.69 dB	23.50 dB	3.498%	0.274	0.0	0.5	0.5
Q5_0	32.36 dB	32.06 dB	2.036%	0.276	0.5	1.0	1.0
Q4_0	30.71 dB	24.70 dB	2.387%	0.296	0.0	1.0	1.0
Q4_1	25.74 dB	22.11 dB	3.293%	0.254	0.0	0.5	0.5
Q5_1	32.98 dB	34.99 dB	1.894%	0.227	0.0	1.0	1.0
Q3_K	25.54 dB	24.72 dB	3.077%	0.288	0.0	0.0	0.5
Q2_K	24.21 dB	19.44 dB	4.447%	0.133	0.0	0.0	0.0

Recommendations:

• Q8_0 — the practical default: 53 dB point-map PSNR and perfect pose within 5 degrees, at 1.33 GiB.
• Q6_K — acceptable if you have headroom in geometry error: 45.7 dB, 0.57% Chamfer, pose still perfect at 5 degrees, but F@1% drops to 0.90.
• Q5_K and below — not viable for production. At Q5_K the point map is already 34.7 dB and AUC@5 is zero; Q4_K collapses to 24.7 dB with a 29-degree mean translation error.
• F16 — near-lossless (71 dB, 0.04% Chamfer) but no VRAM saving.
• Legacy Q4_0/Q4_1/Q5_0/Q5_1 are included for compatibility testing only; none matches Q8_0, and Q4_1/Q3_K/Q2_K lose pose.

These are 3-frame, single-scene numbers. Validate on your own multi-view sequences before deployment; use --dump-npz and the Lux3D compare_geometry.py vggt script.

Usage

--model-path accepts the .gguf file directly or a directory containing model.gguf:

lux3d-cli run vggt --device cuda `
  --model-path models/VGGT-1B-Q8_0.gguf `
  --source frames/ --output out.ply

For tensor-level evaluation, add --dump-npz out.npz. The output point cloud is confidence-filtered; raw world points, depth, and camera tensors are in the NPZ.

Metadata

• general.architecture = vggt
• lux3d.family = vggt
• lux3d.quantization, lux3d.fallback_dtype, lux3d.tensor_count
• lux3d.resolved_config — embedded JSON (embed_dim, img_size, patch_size).

License

Non-commercial. The VGGT-1B weights are CC BY-NC 4.0. See VGGT-LICENSE-CC-BY-NC-4.0.txt. Attribution: Meta AI Research and the University of Oxford VGG group; paper arXiv:2503.11651.

The upstream repository's code license is included separately as VGGT-LICENSE-CODE.txt for reference. These GGUF files are quantized derivatives of the weights and remain non-commercial; keep both license files with them when redistributing.

Commercial use requires the separately distributed facebook/VGGT-1B-Commercial checkpoint and review of Meta's terms.