MrNeRF has released LichtFeld Studio v0.5.3, the release that finishes the open source Gaussian splatting suite’s rendering rewrite by tearing out OpenGL and the CUDA renderer entirely and standing up Vulkan as the single render path, joined by a new Asset Manager subsystem, RAD file format export with a level of detail viewer, quantized optimizer state for lower training VRAM, and 8K-resolution training.
“Bye bye opengl - hello vulkan” lands the migration, so that as of this build there is no CUDA renderer to fall back to; the Vulkan viewer and a Vulkan splat rasterizer are how LichtFeld draws now. That migration had been on the roadmap for a minute. The v0.5.2 release flagged the eventual Vulkan move, RAD support, and the Asset Manager as the v0.5.3 targets, and v0.5.3 ships all three.
Fastgs vksplat brings up the splat path on Vulkan, the forward module is named lfs_vksplat_fwd, while further Vulkan tooling work and fixes shake out the surrounding regressions, and a macro tile inference pipeline for the Vulkan viewer restructures how the viewer evaluates splats.
The migration reads as renderer-only. Training in LichtFeld remains CUDA, a multi-GPU fix that matches the Vulkan physical device to “the CUDA device the trainer is using” confirms the trainer is still CUDA-based, and that shared buffer is what connects the CUDA trainer to the Vulkan viewer, with no indication the trainer itself was moved to Vulkan.
The project has not been alone in betting on Vulkan for 3DGS. NVIDIA’s vkSplatting, the earlier VulkanSplatting, and SplatStream all explored the same API, but LichtFeld is doing it inside a full training and editing suite rather than a standalone renderer.
On the training side the VRAM story is the other reason to upgrade. A quantized Adam optimizer state reduces the memory the optimizer holds during training, and a VRAM/performance HUD lets you watch that number and frame timing live instead of inferring them from nvidia-smi, while a fix for a training speed slowdown closes out a regression that was dragging throughput. Separately, 8K resolution image training arrives on the 3DGS path, specifically the non 3DGUT path.
The Asset Manager is effectively a new subsystem, contributed by shadygm across “Asset Manager” and a run of follow ups. Before this, LichtFeld had no direct place to browse, import, and inspect the splats and assets a session pulls in. The Asset Manager closes that gap and gives the plugin ecosystem a surface to load into on startup. Alongside it, RAD lands as a net new export route. RAD here is the SparkJS Radiance Field LOD file format that World Labs shipped Spark 2.0. LichtFeld authoring its own RAD export and LOD viewer means scenes trained here can target that streaming friendly LOD pipeline directly.
The rest of the changelog fills in the data and editing surfaces. On data loading, a new SPHERICAL (equirectangular) camera model in the COLMAP loader, with EQUIRECTANGULAR settled as the canonical 360 model name, lets the COLMAP path ingest equirectangular captures, animated PLY sequence playback arrives, and fixes land for 3DGS holes, mesh2splat double gamma washed out color, and masks, with in-memory masks and float-dtype colors exposed to direct-scene plugins.
On editing and UI, v0.5.3 adds a color eyedropper for splats, custom gizmos, a reactive UI pass, an RmlUi editor and Tree-sitter integration, GPU instanced camera frustums in the viewport, scene graph groups, selection by distance from world center, CSV/XML/LAS-LAZ file dialogs, a .webp image filter, and a clipboard image read API for plugins.
The Volinga integration also deepens, with a TCP server to broadcast and share events and a new mask mode combining ignore and segment. New contributors this cycle include selimsandal, Vilour, VisibleReality, parodiniccolo, Ediolot (Volinga), iliesaya, and zalo.
LichtFeld Studio remains one of the strongest offerings for gaussian splatting out there and continues to march towards its V0.6 release. You can build LichtFeld Studio from source for free or purchase the one click binary for $30 directly from their website.
