Local Training 101: Postshot, Brush and Lichtfeld Studio

Postshot is one of the most accessible local trainers for non-engineers. It wraps COLMAP alignment and 3DGS training in a single GUI — drag in an image folder to launch the full pipeline. Typical small scenes finish in 10-20 minutes on an RTX 3090 or better. Postshot has free and paid tiers; the free tier limits image count and export formats — upgrade after validating your workflow. Under 8 GB VRAM you may hit OOM; try reducing image resolution or total frame count.

Brush, developed by Arthur Brussee, is one of the most active open-source 3DGS trainers in the community. Built on Rust + WGPU, it runs cross-platform (Windows/macOS/Linux) without hard CUDA dependency, so it can run on non-NVIDIA GPUs (with varying performance). A web demo lets you evaluate it instantly in a browser. The local version supports custom training parameters, live loss curves and splat-count graphs — ideal for users who want to understand the training process. As an active open-source project, features iterate quickly; check Release Notes before use.

Lichtfeld Studio sits between research and production, offering finer-grained parameter control and experimental training strategies. The interface is more complex than Postshot, with a steeper learning curve, but is friendlier for users who need precise hyperparameter tuning or A/B comparisons. It supports multi-scene batch processing, suitable for teams doing scaled Gaussian asset production. Like Brush, use it alongside official documentation; some features require a working knowledge of 3DGS internals. [unverified]

The two most important training log curves are total Loss and Splat Count. Loss should drop steeply in the first few thousand steps, then plateau; persistent oscillation or non-convergence usually signals poor SfM alignment or excessively blurry images. Splat count grows continuously during the adaptive density control phase (~0-15,000 steps) and stabilizes at a level matching scene complexity (tens of thousands to millions for small objects, potentially hundreds of millions for large scenes). Abnormally high splat counts (exceeding VRAM) or abnormally low counts (insufficient detail) should be addressed by improving capture quality, not by tuning training parameters.