What a week for gaussian splatting and it's only Wednesday. MrNeRF has been pushing out nightly builds adding incremental features since mid December and the v0.4.0 in January, but he finally has just pushed out the newest stable version 0.5 and with it several exciting announcements.
The release introduces a redesigned interface, a built-in plugin architecture powered by embedded Python, and several new capabilities across training, editing, and geometry workflows. One of the most visible changes arrives in the interface layer. The project has transitioned its GUI framework away from ImGui toward RmlUI, enabling a styling model built around familiar HTML and CSS concepts.
Version 0.5 also embeds a full Python runtime directly into LichtFeld Studio, bringing what are effectively installable plugins. Each plugin runs inside its own isolated uv environment, ensuring that dependencies remain contained and do not interfere with other extensions. Plugins can be installed with a single click from an integrated marketplace and support hot-reloading, meaning developers can modify and test extensions without restarting the application. Crucially, plugins have access to much of the internal application state. This allows them to interact with scenes, training processes, and geometry data. This alone is a massive update that I am extremely excited to utilize, reflecting the broader momentum behind LichtFeld Studio, which is mainly supported by Core11 GmbH.
Six plugins are available at launch, including integrations for Sharp4D, tools for generating custom depth maps, and a densification workflow tied to the project’s extended bounty program. Because the system is fully programmable, these plugins can implement new training methods, editing tools, or entire reconstruction workflows.
The plugin architecture also integrates directly with the application’s editing history. LichtFeld Studio now includes a full undo and redo stack, and plugin actions can register within this history system so that their operations behave like native features.
The release introduces several improvements to the underlying reconstruction pipeline. Among them is support for ImprovedGS+, a recently published research implementation that focuses on high performance training for LFS. Additionally, NVIDIA's PPISP has been integrated, which models real camera behaviors such as exposure drift, vignetting, white balance, and response curves as physically motivated corrections downstream of the renderer, preventing the scene representation from absorbing photometric inconsistencies. The update also expands the platform’s geometry capabilities. Mesh rendering is now supported alongside Gaussian splats, and Python bindings for OpenMesh have been integrated into the environment. These additions make it easier to work with traditional polygonal geometry within the same workspace as radiance-field representations.
You may also remember the library Mesh2Splat from EA, which converts traditional 3D assets to splats near instantly. Mesh2Splat has been brought to Lichtfeld Studio. Another notable addition is support for the Model Context Protocol (MCP). MCP allows external tools and agents to interact with LichtFeld Studio using the same internal code paths available to the graphical interface. This means automated systems can access the application with capabilities comparable to a human operator, opening the door to programmatic workflows that automate reconstruction, editing, or dataset preparation.
According to sources, the next iteration of LFS will shift focus toward improving training workflows and overall reconstruction quality.
LichtFeld studio remains both open sourced and licensed with a GPL-3.0 license. More information about the platform can be found here.
