LichtFeld-Studio Releases RoMa v1 Weights for Image Matching

The tool LichtFeld-Studio has a new release — RoMa v1 dense matcher weights. This is a permanent model asset in the native .lfw format that is used in the densification pipeline for matching image pairs. The author marked the release as pre-release only so that it would not appear in the Latest badge: this is not an application release, but a model file. The model is assembled from two checkpoints: the matcher Parskatt/RoMa roma_outdoor.pth under the MIT license and the facebookresearch/dinov2 ViT-L/14 backbone under Apache-2.0. It also uses a fine-feature stem from torchvision vgg19_bn under BSD-3-Clause. The file contains 636 tensors, 416,930,219 parameters in float16, and positional embeddings baked for the working resolutions 448 and 560. The file header records the SHA-256 of both source checkpoints, so the asset can be traced back to the exact inputs.

What it means

For LichtFeld-Studio, this is an addition to the scene densification pipeline: the RoMa v1 weights are needed for matching image pairs during densification. The tool itself is a 3D generator based on 3D Gaussian Splatting, open source under GPL-3.0, written in C++23 and requiring CUDA 12.8+ on NVIDIA hardware. Export is supported to PLY, SOG, SPZ and a standalone HTML viewer. This model asset was not previously listed in the tool's fact sheet: the section of the AI for 3D reference describes LichtFeld-Studio as an application for training, inspecting and exporting Gaussian scenes, without mentioning specific matcher weights. Now the tool has a separate downloadable component with a transparent provenance chain: the SHA-256 of the source checkpoints is recorded right in the file header. The release concerns those who use the LichtFeld-Studio densification pipeline: the RoMa v1 weights are a permanent model asset that does not change from version to version of the application.