Skip to main content
Version: 3.4 (unreleased)

AI & Machine Learning Requirements

VRGS includes a set of optional AI / machine-learning features. They are not needed for everyday viewing and interpretation, but when you use them they have their own hardware, disk, and connectivity implications. This page covers what you need.

Features that use AI​

  • Image segmentation (SAM) — interactive "Segment Anything" tools in the photograph view for masking and labelling.
  • Fracture-trace detection (U-Net) — automatic detection of fracture traces on photographs.
  • Dense reconstruction — neural dense matching used in Structure-from-Motion (SfM) workflows.
  • Classifiers / neural networks — additional analysis and classification tools.

GPU acceleration​

The AI models run through ONNX Runtime, which can use either your CPU or an NVIDIA GPU:

  • NVIDIA GPU + CUDA gives a large speed-up and is strongly recommended for AI work. VRGS ships with the supporting CUDA runtime, so a separate CUDA toolkit install is not required — you only need a compatible NVIDIA GPU and an up-to-date driver.
  • CPU fallback — most AI features fall back to the CPU automatically if no compatible GPU is present. They still work, but inference is significantly slower.
  • One exception: the newest SAM3 segmentation model requires an NVIDIA CUDA GPU and will not run on the CPU. The earlier SAM 2.1 models and the fracture-detection U-Net run on either CPU or GPU.

If you intend to use AI features regularly, a dedicated NVIDIA GPU is the single most important choice.

VRAM guidance​

The segmentation models come in several sizes. Larger models are more capable but need more GPU memory:

  • Smaller variants (e.g. SAM 2.1 tiny / small) are suited to GPUs with limited VRAM.
  • Larger variants (e.g. SAM 2.1 large, and SAM3) benefit from more VRAM.

As a rule of thumb: more VRAM lets you run the larger, more accurate models and process bigger images faster. Exact VRAM minimums depend on the model and image size; if a model is too large for your GPU, choose a smaller variant.

Model download & storage​

AI models are not bundled with the installer — they are downloaded on demand the first time you use a feature.

  • The first download requires an internet connection and a GeoTour sign-in (used to authorise the download).

  • Models are cached locally so they only download once, under:

    %LOCALAPPDATA%\VRGeoscience\models

  • Budget a few GB of free disk space for the model cache. The larger segmentation models are several hundred MB to over a GB each.

Once downloaded and cached, the models load locally and no longer need an internet connection.

The AI assistant (Athos / LLM)​

VRGS's built-in AI assistant is cloud-based and does not use your local GPU:

  • It calls hosted LLM services (such as OpenAI or Anthropic Claude), so it needs an internet connection and a valid API key.
  • Optionally, it can connect to a local Ollama server if you prefer to run language models on your own machine; in that case the hardware requirements are determined by Ollama, not VRGS.