MoGe-3 for JAX / Equinox (moge-eqx)

Converted weights of MoGe-3 (Microsoft, MIT licence), a monocular geometry model that predicts point maps, depth, normals and camera intrinsics from a single image, for the JAX/Equinox port moge_eqx.

file	upstream checkpoint
moge3_vitl.eqx	Ruicheng/moge-3-vitl
moge3_vitg.eqx	Ruicheng/moge-3-vitg

These are the same parameters as the upstream PyTorch checkpoints, re-serialised in Equinox leaf order (eqx.tree_serialise_leaves) by the port's scripts/convert_weights.py. The trunk (encoder, neck, the four heads, focal/shift recovery) is ported; the optional sparse refiner is not, and its tensors are not included.

Parity. Against the PyTorch reference, every seam of the trunk scores cosine 1.00000000 with a norm ratio within 4e-06 of 1 (moge3-vitl) and within 2.7e-07 of 1 (moge3-vitg); the recovered focal agrees with upstream's solver to a median 2.4e-07 relative over 20 real images.

Use

from moge_eqx import MoGeModel

model = MoGeModel.from_pretrained("moge3-vitl")   # downloads moge3_vitl.eqx from this repo

from_pretrained reads a local weights/ directory first when the package is a source checkout, then this repository.

Licence and attribution

MoGe and its weights are released by Microsoft under the MIT licence; this conversion is redistributed under the same licence. Please cite the MoGe papers when using these weights: https://github.com/microsoft/MoGe