MoGe-3 (VITL) — MLX
MLX port of MoGe-3: fine-detail monocular geometry estimation with self-guided sparse volumetric refinement. Predicts metric point maps, depth maps, normal maps, valid-pixel masks, and camera intrinsics from a single image.
Converted with mlx-vlm from the official
model.pt checkpoint (fp32).
Usage
from mlx_vlm import load
from mlx_vlm.models.moge3.generate import MoGe3Predictor, read_image
model, processor = load("mlx-community/moge-3-vitl-mlx-fp32")
predictor = MoGe3Predictor(model, processor)
output = predictor.infer(read_image("image.jpg"), resolution_level=9)
points, depth, mask = output["points"], output["depth"], output["mask"]
normal, intrinsics = output["normal"], output["intrinsics"]
Validated against the torch reference (CPU, fp32): identical valid-pixel masks and <0.2% median relative depth error.