LichtFeld-Studio releases MoGe-3 ViT-L weights in native lfw format

LichtFeld-Studio has released a build with MoGe-3 ViT-L base weights in the native .lfw format. This is the default depth model for the SLAM pipeline and the preprocess CLI. The release is marked as pre-release only so that it does not get the Latest badge: it is a permanent model asset, not an application release. Details are on the release page. The model was converted from the microsoft/MoGe checkpoint (MIT, © Microsoft Corporation) and the Ruicheng/moge-3-vitl weights (MIT). The DINOv2 ViT-L/14 backbone components are taken from facebookresearch/dinov2 under Apache-2.0. Only the base model was kept from the original checkpoint: the unused sparse refiner tensors are omitted. The result is 520 tensors, float16, native execution at 1,800 tokens.

What it means

For LichtFeld-Studio, this is a model asset that plugs into the existing pipeline. The tool itself is an open source 3D generator under GPL-3.0, written in C++ and running on CUDA 12.8+. The authors describe it as a native application for training, inspecting, editing, automating and exporting 3D Gaussian Splatting scenes. Export is supported to PLY, SOG, SPZ and a standalone HTML viewer. The tool fact sheet does not describe the MoGe-3 depth model separately: there is no information on VRAM requirements, the size of the weights file in gigabytes, or the supported input image resolutions. The developer does not specify these parameters in the release description. The licensing picture for the asset is composite: the MoGe code and weights are MIT, the DINOv2 components are Apache-2.0. Both licenses permit use and modification, and Apache-2.0 additionally contains an explicit patent grant. The release addresses the need for native integration of the depth model into the LichtFeld-Studio SLAM pipeline without external conversions.