Sumitc13/moge-3-vitg-safetensors

MoGe-3 ViT-G, safetensors

The MoGe-3 ViT-G checkpoint (Ruicheng/moge-3-vitg, Microsoft, MIT) converted 1:1 to safetensors. Same weights, same config, no quantisation; just a format that loads by memory-mapping instead of torch.load, which is faster and needs no pickle.

Made for MoGe-nuke, a Nuke OFX node that runs the full model (sparse refiner included) live.

Files

file	size	sha256
model.safetensors	5.00 GB	685c5bc2bc1acfac86b928255c2c5397a7de4824870ade392e2ba1f74c2ce52b

Source: Ruicheng/moge-3-vitg/model.pt, sha256 ce7c15417e9105c2ace7b4272e2cc69e36940921211eb7fa05d4d0bb03f0a00c. The safetensors metadata header carries model_config (the constructor arguments for moge.model.v3.MoGeModel), source, source_sha256 and moge_version so the file is self-describing.

Loading

import json, torch
from safetensors import safe_open
from safetensors.torch import load_file
from moge.model.v3 import MoGeModel        # pip install git+https://github.com/microsoft/MoGe

path = "model.safetensors"
with safe_open(path, framework="pt") as f:
    config = json.loads(f.metadata()["model_config"])
model = MoGeModel(**config)
model.load_state_dict(load_file(path), strict=True)
model = model.cuda().eval()

out = model.infer(image_tensor)   # (3, H, W) in [0, 1], sRGB-encoded
# out["depth"], out["normal"], out["mask"], out["points"], out["intrinsics"]

Or with the helper shipped in MoGe-nuke:

from moge_safetensors import load_moge     # MoGe-nuke/tools
model = load_moge("model.safetensors", "cuda")

Conversion

python tools/convert_to_safetensors.py model.pt model.safetensors

from the MoGe-nuke repo; the script records the source hash and config in the header, and every tensor is compared to the original after writing.

Licence

MIT, as the original: Copyright (c) Microsoft Corporation. See the MoGe repository.