klartraum.ai

… wake up, dreamer …

Real-time neural rendering and inference on any device with Vulkan, from embedded devices to high-end data center GPUs.

Klartraum Engine

C++ · Vulkan

A compute and rendering engine built on Vulkan. Work is described as a graph of GPU operations that is recorded into command buffers once and then only submitted, so the CPU stays out of the way for highest performance.

Goals

  • Run neural networks on any Vulkan device
  • Combine Gaussian splatting, neural networks and classic rasterization in one pipeline
  • Pre-record all GPU work for low latency and low CPU load
  • Work in a window or headless, as an inference engine

Klartraum Studio

Node editor · live preview

Building, inspecting and tuning Klartraum compute graphs. Nodes stand for the engine's compute graph building blocks, and edits are compiled and shown live.

Goals

  • Build and inspect compute graphs visually
  • See results live while editing
  • Run one-off graphs for previews and image files
  • Use Gaussian splatting, ONNX models and tensor operations as nodes