MoGe-3 ViT-L — MLX (fp32)
Microsoft's MoGe-3 monocular geometry model converted to MLX safetensors for
Apple Silicon. 607 tensors, float32.
Usage
from mlx_vlm import load
from mlx_vlm.extraction import extract
model, processor = load("nativ-community/moge-3-vitl-fp32")
result = extract(model, processor, "photo.jpg")
print(result["depth"].shape, result["normal"].shape, result["intrinsics"])
From the command line:
mlx_vlm.extract --model nativ-community/moge-3-vitl-fp32 \
--image photo.jpg --output geometry.npz
Named outputs are points, depth, normal, mask and intrinsics. Masked-out
pixels stay infinite, so read geometry through mask.
Conversion
python -m mlx_vlm.models.moge3.convert \
--torch-ckpt model.pt --mlx-path moge-3-vitl-fp32 --dtype float32
model.pt comes from Ruicheng/moge-3-vitl. Conversion needs torch;
inference does not.
Verification
On a 640x480 photo this checkpoint returns depth (480, 640),
normal (480, 640, 3), a boolean mask and 3x3 intrinsics.
Attribution
Weights derive from Ruicheng/moge-3-vitl, MIT licensed, redistributed here
converted to MLX under the same licence.
