akaiii/meadow-world-builder-weights

Meadow - World Builder · Pretrained Weights

MLX-converted weights for single-image → 3D Gaussian Splatting inference on Apple Silicon.

Code: https://github.com/Hey-Meadow/meadow-world-builder

Quickstart

# 1. Clone the code repo + install
git clone https://github.com/Hey-Meadow/meadow-world-builder
cd meadow-world-builder && python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt && pip install -e .

# 2. Download these weights
huggingface-cli download akaiii/meadow-world-builder-weights \
    --local-dir meadow_wb/weights/sam3d_objects

# 3. Run inference
python meadow_wb/infer.py --image YOUR.png --mask YOUR_MASK.png \
    --use-moge --use-shortcut --dtype mixed --prune-outliers \
    --out output.ply

Files

File	Size	Module
ss_flow.npz	3.3 GB	Sparse-Structure DiT
slat_flow.npz	2.0 GB	SLAT DiT
ss_embedder.npz	2.2 GB	image conditioning (SS)
slat_embedder.npz	2.2 GB	image conditioning (SLAT)
ss_decoder.npz	166 MB	SS occupancy decoder
slat_decoder_gs_4.npz	192 MB	Gaussian decoder (4 splats / voxel)
slat_decoder_gs.npz	192 MB	Gaussian decoder (8 splats / voxel)
moge_vitl.npz	1.2 GB	MoGe ViT-L depth backbone

Total ≈ 11.5 GB on disk.

Performance

Object	M1 Max wall-clock	Output
chair	86 s	63 624 Gaussians
table	94 s	64 000 Gaussians
plush	122 s	64 000 Gaussians

See benchmark report.

License

These weights are derivative MLX conversions of upstream pretrained components. Use is governed by the bundled LICENSE and the upstream licences referenced therein.