Asher-1/querysplat_vggto

QuerySplat GGUF (pure C++ ggml inference)

Pre-converted QuerySplat weights for the pure C++ (ggml) inference chain in querysplat-ggml: a feed-forward pipeline that turns a sparse multi-view image set into 3D Gaussian splats with zero PyTorch at runtime. The VGGT-Omega encoder GGUFs live in the companion repo Asher-1/map_anything_vggt (vggt-omega-1b-512-{f32,f16,q8_0}.gguf — the only encoder family the QuerySplat decoders accept).

images ─► [encoder GGUF: VGGT-Omega-1B-512] ─► pose + depth + cached-layer tokens
                                                       │
             [decoder GGUF: QuerySplat-paper] ◄────────┘
                       │
              524,288 gaussians ─► rasterizer ─► renders

Files

File	Size	Role	sha256
querysplat-paper-f32.gguf	4.47 GB	decoder, parity tier (default)	e59558d4cda480b743788a56eb5bb27d07fadb42be87e852ff9868002c07e188
querysplat-paper-f16.gguf	2.30 GB	decoder, compact tier (bit-identical render — weights staged as F32)	e26b237544f0b803df7311131a8ecfa999fed79f4f6c56f06d0748dba6872905
querysplat-paper-q8_0.gguf	1.31 GB	decoder, smallest tier (cpp-vs-GT within 0.04 dB of f32)	aaeaa570ce59edfa5cd725ea4ac33cc6a20e95966b1daa1d31922eef627f5f38
lpips-vgg-f32.gguf	56 MB	LPIPS-vgg metric weights for querysplat-eval (never quantized: the metric is the judge)	0a2c7475ad0eed64cd40ec7b221b3dc639cc7ed142176554a30cd07aac1c7b7d

Encoder (from Asher-1/map_anything_vggt): vggt-omega-1b-512-f32.gguf (4.26 GB, parity) · vggt-omega-1b-512-f16.gguf (2.13 GB) · vggt-omega-1b-512-q8_0.gguf (1.23 GB). The encoder and decoder of a run must use the same quant tier.

Usage

# one command (builds if needed, downloads nothing else)
./run_querysplat.sh cpp --input your/images --out outputs/scene

# or directly, official infer.py layout out of the box
cpp_ggml/build-*/bin/querysplat-cli \
    --encoder  vggt-omega-1b-512-f32.gguf \
    --decoder  querysplat-paper-f32.gguf \
    --input_folder your/images --output_dir outputs/scene

Backends: CUDA / Vulkan / CPU (compile flags -DMAP_GGML_CUDA=ON / VULKAN). The f32-accuracy kernels (exact-F32 attention + cuBLAS, matching torch) are default-ON and are explicit options in the API — no environment variables.

Measured (RTX 4090 + i9-14900K, S=4, 512×512)

Row	enc+dec s	render s	e2e s	vs official py CUDA
official py (torch f32) CUDA	2.59	0.30	2.89	1.00×
cpp Vulkan q8_0	2.50	0.39	2.90	1.00×
cpp CUDA f32	2.81	0.40	3.20	1.11×
cpp CPU f32	84.9	0.40	85.3	29.5×

Quality (mean over 3 backends × 3 scenes): cpp-vs-py render agreement 59.2 dB (f32) / 45.6 dB (f16) / 34.5 dB (q8_0); cpp-vs-GT within 0.05 dB of the official torch chain on every tier (26.84–26.89 dB). Full methodology, per-seam parity gates and the DL3DV protocol table: MODEL_CARDS.md · querysplat_report.md.

Provenance

• Decoder: inspatio/querysplat querysplat_vggto_1B_512_paper.safetensors (sh0), converted with cpp_ggml/scripts/convert_querysplat_to_gguf.py (tensor names shortened to ggml's 64-char key limit; gguf-py shape reversal per ggml ne convention).
• Encoder: facebook/VGGT-Omega vggt_omega_1b_512.pt, converted with the map-anything-ggml convert_vggt_omega_to_gguf.py.
• LPIPS: torchvision VGG16 features + lpips v0.1 linear heads, converted with cpp_ggml/scripts/convert_lpips_to_gguf.py.

The QuerySplat decoder weights inherit the upstream inspatio/querysplat release terms; the conversion tooling and the inference chain are Apache-2.0.

Citation

@article{querysplat2025,
  title  = {QuerySplat: Query-Based 3D Gaussian Splatting for Sparse-View Reconstruction},
  author = {inspatio},
  note   = {\url{https://github.com/inspatio/querysplat}}
}