Asher-1/SAM_3D_OBJECTS_GGUF

github: https://github.com/Asher-1/sam-3d-objects-ggml

SAM 3D Objects GGUF Models

All GGUF weights belong in gguf/, never in benchmarks. Native raw reconstruction consumes the MoGe model plus five generative stages and exports a Gaussian PLY, a textured GLB and its base-color PNG. Vulkan inference hands its raw mesh and Gaussian outputs to a separate CUDA PBR process.

Download and conversion

The matching model repository is Asher-1/SAM_3D_OBJECTS_GGUF. Weights inherit the upstream SAM 3D Objects license. Use the helper from the checkout root; it refuses to overwrite an existing local model by default:

bash cpp_ggml/scripts/download_gguf.sh --help
bash cpp_ggml/scripts/download_gguf.sh --dtype q8_0
SAM3D_PYTHON=/path/to/sam3d-objects/bin/python \
  bash cpp_ggml/scripts/prepare_moge_gguf.sh

The published MoGe ViT-L file is the F16 GGUF (moge_vitl-f16.gguf), also in the public model repository and part of the downloader's default f16 set. The helper still reconverts from the official Ruicheng/moge-vitl checkpoint (huggingface_hub) when another dtype or a fresh validation against upstream is needed. It preserves existing GGUF files unless explicitly given --force.

For conversion from local official checkpoints (conversion writes the chosen output filenames; do not overwrite a benchmark baseline unintentionally):

SAM3D_PYTHON=/path/to/sam3d-objects/bin/python
"$SAM3D_PYTHON" cpp_ggml/scripts/convert_sam3d_to_gguf.py \
  --checkpoint-dir checkpoints/hf --model all --dtype q8_0 \
  --keep-f16-regex '^cemb\.' \
  --output cpp_ggml/models/gguf

Repeat with --dtype f16 or --dtype q8_0 for the other full-matrix families. --model all converts the six generative stages, not MoGe; download the published F16 file with the downloader or prepare it with the separate helper above. The checkpoint directory must contain the converter's expected stage files; see the runtime guide. Do not substitute LingBot-Map GGUF weights: that is a different architecture.

Current local inventory

Sizes below are actual local file sizes after the 2026-09-14 asset ruling, in MiB (2^20 bytes). They are disk sizes, not inference VRAM requirements or a promise that a model downloaded under the same filename has identical contents.

Stage	F16 MiB	F32 MiB	Q4_K MiB	Q8_0 MiB	Role
MoGe ViT-L	600.0	removed	not distributed	not distributed	image/mask point-map conditioning
SS generator	3052.4	removed	864.8	2191.6	condition encoder, structure diffusion and pose
SS decoder	140.6	removed	39.7	74.8	sparse support
SLat generator	2341.9	removed	663.9	1247.5	structured latent diffusion
Gaussian decoder	163.7	removed	47.2	87.7	Gaussian attributes
Gaussian stride-4 decoder	162.4	removed	45.9	86.4	optional alternative, not the current raw matrix
Mesh decoder	173.5	removed	49.1	92.4	FlexiCubes mesh features

The 2026-09-14 ruling deleted every Q4_0/Q4_1 GGUF (12 files, 3.6 GB) and moge_vitl-f32.gguf after measured A/B arbitration (see "Weight-Asset Rulings" in the parity contract): q4_k is the accuracy/speed winner inside the Q4 family and matches the historical q4_best selection; the F16-weight MoGe matches the official F32 checkpoint to MAE 2.6e-4 - better than the F32-weight GGUF - at half the size. ss_generator-q8_0.gguf was re-exported with --keep-f16-regex '^cemb\.' so the DINO/PointPatch/fuser weights stay F16 regardless of the deployment dtype (hence 2191.6 MiB, not 1625.7). convert_sam3d_to_gguf.py keeps q4_0 in its choices only for the QAT retraining pipeline; q4_1 was removed from it.

Precision policy

F16, Q8_0 and Q4_K select the five generative files of that suffix; the raw matrix always uses the same explicitly recorded MoGe F16 file. Norm/bias and other tensors unsupported by a quantization layout retain the converter's floating-point types. Execution uses mixed intermediate dtypes, not uniformly F32 activations. The actual graph is recorded by --dtype-contract-out. The historical sensitive-layer SS Q4_K candidate provenance remains in q4_best. No GGUF is moved, replaced or retrained when publishing benchmarks.

End-to-end verification

Current per-model latency, final GLB renders, standalone textures and explicit gate results are linked from benchmarks. Only these complete raw-input results represent release evidence. A conditioned-stage replay, a load test or an improved Q4 trajectory does not establish complete reconstruction parity.

The root run_ggml.sh launcher uses strict SS attention. Full output matches the official material scope: baked base color with a PBR material, not independently inferred metallic/roughness/normal maps. Native CUDA or Vulkan-to-CUDA reconstruction does not require Torch or cuDNN at runtime. VTK >= 9.3, OpenCV and the explicit native rasterizer/MeshFix license options are required for complete post-processing.

VAE encoders are training components and are not required by generation.