LichtFeld-Studio releases SAM 2.1 Hiera base+ video weights in lfw format

On October 28, 2026, a release with SAM 2.1 Hiera base+ weights for video was published in the LichtFeld-Studio repository. This is a permanent model asset, not an application release: it is marked as pre-release only so that it does not appear in the Latest badge. The weights include the video components — image encoder, prompt and mask decoders, temporal memory and object pointers — and are saved in the native .lfw format used by the built-in video masking. The source is the official sam2.1_hiera_base_plus.pt checkpoint from Meta, license Apache-2.0. Details are in the release description on GitHub.

What it means

This news concerns LichtFeld-Studio, an open source 3D generator based on Gaussian Splatting by developer janusch. The code is distributed under GPL-3.0, written in C++ and designed for CUDA 12.8+ with NVIDIA hardware. The application can train, inspect, edit and export 3D scenes, including to PLY. Our tool fact sheet does not describe video masking or the .lfw format: the tool page records export to PLY, SOG, SPZ and an HTML viewer, as well as the latest release v0.5.3. The new asset adds support for the SAM 2.1 video components to LichtFeld-Studio, but the developer does not specify the model's parameters in the release description — how much it weighs, how much VRAM it requires and exactly how it is launched. All that is known is that the conversion is reproducible with the export_sam2_weights.py script with the --video flag, and that after it 615 tensors and 80 850 434 Float32 values are identical to the original checkpoint. For those who use LichtFeld-Studio for 3D scenes, this release is interesting primarily as an extension of native video masking. Other readers choosing tools for working with video and neural networks should look at the AI for video section of the reference — it collects current options for different tasks and hardware.